notaz.gp2x.de / ia32rtools.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
minor fixes
[ia32rtools.git] / tools / translate.c
... / ...
CommitLineData
1/*
2 * ia32rtools
3 * (C) notaz, 2013-2015
4 *
5 * This work is licensed under the terms of 3-clause BSD license.
6 * See COPYING file in the top-level directory.
7 *
8 * recognized asm hint comments:
9 * sctattr - function attributes (see code)
10 * sctend - force end of function/chunk
11 * sctpatch: <p> - replace current asm line with <p>
12 * sctproto: <p> - prototype of ref'd function or struct
13 * sctref - variable is referenced, make global
14 * sctskip_start - start of skipped code chunk (inclusive)
15 * sctskip_end - end of skipped code chunk (inclusive)
16 */
17
18#define _GNU_SOURCE
19#include <stdio.h>
20#include <stdlib.h>
21#include <stddef.h>
22#include <string.h>
23#include <math.h>
24#include <errno.h>
25
26#include "my_assert.h"
27#include "my_str.h"
28#include "common.h"
29
30#include "protoparse.h"
31
32static const char *asmfn;
33static int asmln;
34static FILE *g_fhdr;
35
36#define anote(fmt, ...) \
37 printf("%s:%d: note: " fmt, asmfn, asmln, ##__VA_ARGS__)
38#define awarn(fmt, ...) \
39 printf("%s:%d: warning: " fmt, asmfn, asmln, ##__VA_ARGS__)
40#define aerr(fmt, ...) do { \
41 printf("%s:%d: error: " fmt, asmfn, asmln, ##__VA_ARGS__); \
42 fcloseall(); \
43 exit(1); \
44} while (0)
45
46#include "masm_tools.h"
47
48enum op_flags {
49 OPF_RMD = (1 << 0), /* removed from code generation */
50 OPF_DATA = (1 << 1), /* data processing - writes to dst opr */
51 OPF_FLAGS = (1 << 2), /* sets flags */
52 OPF_JMP = (1 << 3), /* branch, call */
53 OPF_CJMP = (1 << 4), /* cond. branch (cc or jecxz/loop) */
54 OPF_CC = (1 << 5), /* uses flags */
55 OPF_TAIL = (1 << 6), /* ret or tail call */
56 OPF_RSAVE = (1 << 7), /* push/pop is local reg save/load */
57 OPF_REP = (1 << 8), /* prefixed by rep */
58 OPF_REPZ = (1 << 9), /* rep is repe/repz */
59 OPF_REPNZ = (1 << 10), /* rep is repne/repnz */
60 OPF_FARG = (1 << 11), /* push collected as func arg */
61 OPF_FARGNR = (1 << 12), /* push collected as func arg (no reuse) */
62 OPF_EBP_S = (1 << 13), /* ebp used as scratch here, not BP */
63 OPF_DF = (1 << 14), /* DF flag set */
64 OPF_ATAIL = (1 << 15), /* tail call with reused arg frame */
65 OPF_32BIT = (1 << 16), /* enough to do 32bit for this op */
66 OPF_LOCK = (1 << 17), /* op has lock prefix */
67 OPF_VAPUSH = (1 << 18), /* vararg ptr push (as call arg) */
68 OPF_DONE = (1 << 19), /* already fully handled by analysis */
69 OPF_PPUSH = (1 << 20), /* part of complex push-pop graph */
70 OPF_NOREGS = (1 << 21), /* don't track regs of this op */
71 OPF_FPUSH = (1 << 22), /* pushes x87 stack */
72 OPF_FPOP = (1 << 23), /* pops x87 stack */
73 OPF_FPOPP = (1 << 24), /* pops x87 stack twice */
74 OPF_FSHIFT = (1 << 25), /* x87 stack shift is actually needed */
75 OPF_FINT = (1 << 26), /* integer float op arg */
76};
77
78enum op_op {
79 OP_INVAL,
80 OP_NOP,
81 OP_PUSH,
82 OP_POP,
83 OP_PUSHA,
84 OP_POPA,
85 OP_LEAVE,
86 OP_MOV,
87 OP_LEA,
88 OP_MOVZX,
89 OP_MOVSX,
90 OP_XCHG,
91 OP_NOT,
92 OP_XLAT,
93 OP_CDQ,
94 OP_BSWAP,
95 OP_LODS,
96 OP_STOS,
97 OP_MOVS,
98 OP_CMPS,
99 OP_SCAS,
100 OP_RDTSC,
101 OP_CPUID,
102 OP_STD,
103 OP_CLD,
104 OP_RET,
105 OP_ADD,
106 OP_SUB,
107 OP_AND,
108 OP_OR,
109 OP_XOR,
110 OP_SHL,
111 OP_SHR,
112 OP_SAR,
113 OP_SHLD,
114 OP_SHRD,
115 OP_ROL,
116 OP_ROR,
117 OP_RCL,
118 OP_RCR,
119 OP_ADC,
120 OP_SBB,
121 OP_BSF,
122 OP_BSR,
123 OP_INC,
124 OP_DEC,
125 OP_NEG,
126 OP_MUL,
127 OP_IMUL,
128 OP_DIV,
129 OP_IDIV,
130 OP_TEST,
131 OP_CMP,
132 OP_CALL,
133 OP_JMP,
134 OP_JECXZ,
135 OP_LOOP,
136 OP_JCC,
137 OP_SCC,
138 // x87
139 OP_FLD,
140 OP_FILD,
141 OP_FLDc,
142 OP_FST,
143 OP_FIST,
144 OP_FABS,
145 OP_FADD,
146 OP_FDIV,
147 OP_FMUL,
148 OP_FSUB,
149 OP_FDIVR,
150 OP_FSUBR,
151 OP_FIADD,
152 OP_FIDIV,
153 OP_FIMUL,
154 OP_FISUB,
155 OP_FIDIVR,
156 OP_FISUBR,
157 OP_FCOM,
158 OP_FNSTSW,
159 OP_FCHS,
160 OP_FCOS,
161 OP_FPATAN,
162 OP_FPTAN,
163 OP_FSIN,
164 OP_FSQRT,
165 OP_FXCH,
166 OP_FYL2X,
167 // mmx
168 OP_EMMS,
169 // pseudo-ops for lib calls
170 OPP_ALLSHL,
171 OPP_ALLSHR,
172 OPP_FTOL,
173 OPP_CIPOW,
174 OPP_ABORT,
175 // undefined
176 OP_UD2,
177};
178
179enum opr_type {
180 OPT_UNSPEC,
181 OPT_REG,
182 OPT_REGMEM,
183 OPT_LABEL,
184 OPT_OFFSET,
185 OPT_CONST,
186};
187
188// must be sorted (larger len must be further in enum)
189enum opr_lenmod {
190 OPLM_UNSPEC,
191 OPLM_BYTE,
192 OPLM_WORD,
193 OPLM_DWORD,
194 OPLM_QWORD,
195};
196
197#define MAX_EXITS 128
198
199#define MAX_OPERANDS 3
200#define NAMELEN 112
201
202#define OPR_INIT(type_, lmod_, reg_) \
203 { type_, lmod_, reg_, }
204
205struct parsed_opr {
206 enum opr_type type;
207 enum opr_lenmod lmod;
208 int reg;
209 unsigned int is_ptr:1; // pointer in C
210 unsigned int is_array:1; // array in C
211 unsigned int type_from_var:1; // .. in header, sometimes wrong
212 unsigned int size_mismatch:1; // type override differs from C
213 unsigned int size_lt:1; // type override is larger than C
214 unsigned int segment:7; // had segment override (enum segment)
215 const struct parsed_proto *pp; // for OPT_LABEL
216 unsigned int val;
217 char name[NAMELEN];
218};
219
220struct parsed_op {
221 enum op_op op;
222 struct parsed_opr operand[MAX_OPERANDS];
223 unsigned int flags;
224 unsigned char pfo;
225 unsigned char pfo_inv;
226 unsigned char operand_cnt;
227 unsigned char p_argnum; // arg push: call's saved arg #
228 unsigned char p_arggrp; // arg push: arg group # for above
229 unsigned char p_argpass;// arg push: arg of host func
230 short pad;
231 int regmask_src; // all referensed regs
232 int regmask_dst;
233 int pfomask; // flagop: parsed_flag_op that can't be delayed
234 int cc_scratch; // scratch storage during analysis
235 int bt_i; // branch target for branches
236 struct parsed_data *btj;// branch targets for jumptables
237 struct parsed_proto *pp;// parsed_proto for OP_CALL
238 void *datap;
239 int asmln;
240};
241
242// datap:
243// on start: function/data type hint (sctproto)
244// after analysis:
245// (OPF_CC) - points to one of (OPF_FLAGS) that affects cc op
246// OP_PUSH - points to OP_POP in complex push/pop graph
247// OP_POP - points to OP_PUSH in simple push/pop pair
248// OP_FCOM - needed_status_word_bits | (is_z_check << 16)
249
250struct parsed_equ {
251 char name[64];
252 enum opr_lenmod lmod;
253 int offset;
254};
255
256struct parsed_data {
257 char label[256];
258 enum opr_type type;
259 enum opr_lenmod lmod;
260 int count;
261 int count_alloc;
262 struct {
263 union {
264 char *label;
265 unsigned int val;
266 } u;
267 int bt_i;
268 } *d;
269};
270
271struct label_ref {
272 int i;
273 struct label_ref *next;
274};
275
276enum ida_func_attr {
277 IDAFA_BP_FRAME = (1 << 0),
278 IDAFA_LIB_FUNC = (1 << 1),
279 IDAFA_STATIC = (1 << 2),
280 IDAFA_NORETURN = (1 << 3),
281 IDAFA_THUNK = (1 << 4),
282 IDAFA_FPD = (1 << 5),
283};
284
285// sctattr
286enum sct_func_attr {
287 SCTFA_CLEAR_SF = (1 << 0), // clear stack frame
288 SCTFA_CLEAR_REGS = (1 << 1), // clear registers (mask)
289 SCTFA_RM_REGS = (1 << 2), // don't emit regs (mask)
290 SCTFA_NOWARN = (1 << 3), // don't try to detect problems
291 SCTFA_ARGFRAME = (1 << 4), // copy all args to a struct, in order
292 SCTFA_UA_FLOAT = (1 << 5), // emit float i/o helpers for alignemnt
293};
294
295enum x87_const {
296 X87_CONST_1 = 1,
297 X87_CONST_L2T,
298 X87_CONST_L2E,
299 X87_CONST_PI,
300 X87_CONST_LG2,
301 X87_CONST_LN2,
302 X87_CONST_Z,
303};
304
305enum segment {
306 SEG_CS = 1,
307 SEG_DS,
308 SEG_SS,
309 SEG_ES,
310 SEG_FS,
311 SEG_GS,
312};
313
314#define MAX_OPS 4096
315#define MAX_ARG_GRP 2
316
317static struct parsed_op ops[MAX_OPS];
318static struct parsed_equ *g_eqs;
319static int g_eqcnt;
320static char *g_labels[MAX_OPS];
321static struct label_ref g_label_refs[MAX_OPS];
322static const struct parsed_proto *g_func_pp;
323static struct parsed_data *g_func_pd;
324static int g_func_pd_cnt;
325static int g_func_lmods;
326static char g_func[256];
327static char g_comment[256];
328static int g_bp_frame;
329static int g_sp_frame;
330static int g_stack_frame_used;
331static int g_stack_fsz;
332static int g_seh_found;
333static int g_seh_size;
334static int g_ida_func_attr;
335static int g_sct_func_attr;
336static int g_stack_clear_start; // in dwords
337static int g_stack_clear_len;
338static int g_regmask_init;
339static int g_regmask_rm;
340static int g_skip_func;
341static int g_allow_regfunc;
342static int g_allow_user_icall;
343static int g_nowarn_reguse;
344static int g_quiet_pp;
345static int g_header_mode;
346
347#define ferr(op_, fmt, ...) do { \
348 printf("%s:%d: error %u: [%s] '%s': " fmt, asmfn, (op_)->asmln, \
349 __LINE__, g_func, dump_op(op_), ##__VA_ARGS__); \
350 fcloseall(); \
351 exit(1); \
352} while (0)
353#define fnote(op_, fmt, ...) \
354 printf("%s:%d: note: [%s] '%s': " fmt, asmfn, (op_)->asmln, g_func, \
355 dump_op(op_), ##__VA_ARGS__)
356
357#define ferr_assert(op_, cond) do { \
358 if (!(cond)) ferr(op_, "assertion '%s' failed\n", #cond); \
359} while (0)
360
361#define IS_OP_INDIRECT_CALL(op_) \
362 ((op_)->op == OP_CALL && (op_)->operand[0].type != OPT_LABEL)
363
364const char *regs_r32[] = {
365 "eax", "ebx", "ecx", "edx", "esi", "edi", "ebp", "esp",
366 // not r32, but list here for easy parsing and printing
367 "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",
368 "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)"
369};
370const char *regs_r16[] = { "ax", "bx", "cx", "dx", "si", "di", "bp", "sp" };
371const char *regs_r8l[] = { "al", "bl", "cl", "dl" };
372const char *regs_r8h[] = { "ah", "bh", "ch", "dh" };
373
374enum x86_regs {
375 xUNSPEC = -1,
376 xAX, xBX, xCX, xDX,
377 xSI, xDI, xBP, xSP,
378 xMM0, xMM1, xMM2, xMM3, // mmx
379 xMM4, xMM5, xMM6, xMM7,
380 xST0, xST1, xST2, xST3, // x87
381 xST4, xST5, xST6, xST7,
382};
383
384#define mxAX (1 << xAX)
385#define mxBX (1 << xBX)
386#define mxCX (1 << xCX)
387#define mxDX (1 << xDX)
388#define mxSP (1 << xSP)
389#define mxST0 (1 << xST0)
390#define mxST1 (1 << xST1)
391#define mxST1_0 (mxST1 | mxST0)
392#define mxST7_2 (0xfc << xST0)
393#define mxSTa (0xff << xST0)
394
395// possible basic comparison types (without inversion)
396enum parsed_flag_op {
397 PFO_O, // 0 OF=1
398 PFO_C, // 2 CF=1
399 PFO_Z, // 4 ZF=1
400 PFO_BE, // 6 CF=1||ZF=1
401 PFO_S, // 8 SF=1
402 PFO_P, // a PF=1
403 PFO_L, // c SF!=OF
404 PFO_LE, // e ZF=1||SF!=OF
405};
406
407#define PFOB_O (1 << PFO_O)
408#define PFOB_C (1 << PFO_C)
409#define PFOB_Z (1 << PFO_Z)
410#define PFOB_S (1 << PFO_S)
411
412static const char *parsed_flag_op_names[] = {
413 "o", "c", "z", "be", "s", "p", "l", "le"
414};
415
416static int char_array_i(const char *array[], size_t len, const char *s)
417{
418 int i;
419
420 for (i = 0; i < len; i++)
421 if (IS(s, array[i]))
422 return i;
423
424 return -1;
425}
426
427static void printf_number(char *buf, size_t buf_size,
428 unsigned long number)
429{
430 // output in C-friendly form
431 snprintf(buf, buf_size, number < 10 ? "%lu" : "0x%02lx", number);
432}
433
434static int check_segment_prefix(const char *s)
435{
436 if (s[0] == 0 || s[1] != 's' || s[2] != ':')
437 return 0;
438
439 switch (s[0]) {
440 case 'c': return SEG_CS;
441 case 'd': return SEG_DS;
442 case 's': return SEG_SS;
443 case 'e': return SEG_ES;
444 case 'f': return SEG_FS;
445 case 'g': return SEG_GS;
446 default: return 0;
447 }
448}
449
450static int parse_reg(enum opr_lenmod *reg_lmod, const char *s)
451{
452 int reg;
453
454 reg = char_array_i(regs_r32, ARRAY_SIZE(regs_r32), s);
455 if (reg >= 8) {
456 *reg_lmod = OPLM_QWORD;
457 return reg;
458 }
459 if (reg >= 0) {
460 *reg_lmod = OPLM_DWORD;
461 return reg;
462 }
463 reg = char_array_i(regs_r16, ARRAY_SIZE(regs_r16), s);
464 if (reg >= 0) {
465 *reg_lmod = OPLM_WORD;
466 return reg;
467 }
468 reg = char_array_i(regs_r8h, ARRAY_SIZE(regs_r8h), s);
469 if (reg >= 0) {
470 *reg_lmod = OPLM_BYTE;
471 return reg;
472 }
473 reg = char_array_i(regs_r8l, ARRAY_SIZE(regs_r8l), s);
474 if (reg >= 0) {
475 *reg_lmod = OPLM_BYTE;
476 return reg;
477 }
478
479 return -1;
480}
481
482static int parse_indmode(char *name, int *regmask, int need_c_cvt)
483{
484 enum opr_lenmod lmod;
485 char cvtbuf[256];
486 char *d = cvtbuf;
487 char *s = name;
488 char w[64];
489 long number;
490 int reg;
491 int c = 0;
492
493 *d = 0;
494
495 while (*s != 0) {
496 d += strlen(d);
497 while (my_isblank(*s))
498 s++;
499 for (; my_issep(*s); d++, s++)
500 *d = *s;
501 while (my_isblank(*s))
502 s++;
503 *d = 0;
504
505 // skip '?s:' prefixes
506 if (check_segment_prefix(s))
507 s += 3;
508
509 s = next_idt(w, sizeof(w), s);
510 if (w[0] == 0)
511 break;
512 c++;
513
514 reg = parse_reg(&lmod, w);
515 if (reg >= 0) {
516 *regmask |= 1 << reg;
517 goto pass;
518 }
519
520 if ('0' <= w[0] && w[0] <= '9') {
521 number = parse_number(w, 0);
522 printf_number(d, sizeof(cvtbuf) - (d - cvtbuf), number);
523 continue;
524 }
525
526 // probably some label/identifier - pass
527
528pass:
529 snprintf(d, sizeof(cvtbuf) - (d - cvtbuf), "%s", w);
530 }
531
532 if (need_c_cvt)
533 strcpy(name, cvtbuf);
534
535 return c;
536}
537
538static int is_reg_in_str(const char *s)
539{
540 int i;
541
542 if (strlen(s) < 3 || (s[3] && !my_issep(s[3]) && !my_isblank(s[3])))
543 return 0;
544
545 for (i = 0; i < ARRAY_SIZE(regs_r32); i++)
546 if (!strncmp(s, regs_r32[i], 3))
547 return 1;
548
549 return 0;
550}
551
552static const char *parse_stack_el(const char *name, char *extra_reg,
553 int *base_val, int early_try)
554{
555 const char *p, *p2, *s;
556 char *endp = NULL;
557 char buf[32];
558 long val = -1;
559 int len;
560
561 if (g_bp_frame || early_try)
562 {
563 p = name;
564 if (IS_START(p + 3, "+ebp+") && is_reg_in_str(p)) {
565 p += 4;
566 if (extra_reg != NULL) {
567 strncpy(extra_reg, name, 3);
568 extra_reg[4] = 0;
569 }
570 }
571
572 if (IS_START(p, "ebp+")) {
573 p += 4;
574
575 p2 = strchr(p, '+');
576 if (p2 != NULL && is_reg_in_str(p)) {
577 if (extra_reg != NULL) {
578 strncpy(extra_reg, p, p2 - p);
579 extra_reg[p2 - p] = 0;
580 }
581 p = p2 + 1;
582 }
583
584 if (!('0' <= *p && *p <= '9'))
585 return p;
586
587 return NULL;
588 }
589 }
590
591 if (!IS_START(name, "esp+"))
592 return NULL;
593
594 s = name + 4;
595 p = strchr(s, '+');
596 if (p) {
597 if (is_reg_in_str(s)) {
598 if (extra_reg != NULL) {
599 strncpy(extra_reg, s, p - s);
600 extra_reg[p - s] = 0;
601 }
602 s = p + 1;
603 p = strchr(s, '+');
604 if (p == NULL)
605 aerr("%s IDA stackvar not set?\n", __func__);
606 }
607 if ('0' <= *s && *s <= '9') {
608 if (s[0] == '0' && s[1] == 'x')
609 s += 2;
610 len = p - s;
611 if (len < sizeof(buf) - 1) {
612 strncpy(buf, s, len);
613 buf[len] = 0;
614 errno = 0;
615 val = strtol(buf, &endp, 16);
616 if (val == 0 || *endp != 0 || errno != 0) {
617 aerr("%s num parse fail for '%s'\n", __func__, buf);
618 return NULL;
619 }
620 }
621 p++;
622 }
623 else {
624 // probably something like [esp+arg_4+2]
625 p = s;
626 val = 0;
627 }
628 }
629 else
630 p = name + 4;
631
632 if ('0' <= *p && *p <= '9')
633 return NULL;
634
635 if (base_val != NULL)
636 *base_val = val;
637 return p;
638}
639
640static int guess_lmod_from_name(struct parsed_opr *opr)
641{
642 if (IS_START(opr->name, "dword_") || IS_START(opr->name, "off_")) {
643 opr->lmod = OPLM_DWORD;
644 return 1;
645 }
646 if (IS_START(opr->name, "word_")) {
647 opr->lmod = OPLM_WORD;
648 return 1;
649 }
650 if (IS_START(opr->name, "byte_")) {
651 opr->lmod = OPLM_BYTE;
652 return 1;
653 }
654 if (IS_START(opr->name, "qword_")) {
655 opr->lmod = OPLM_QWORD;
656 return 1;
657 }
658 return 0;
659}
660
661static int guess_lmod_from_c_type(enum opr_lenmod *lmod,
662 const struct parsed_type *c_type)
663{
664 static const char *qword_types[] = {
665 "uint64_t", "int64_t", "__int64",
666 };
667 static const char *dword_types[] = {
668 "uint32_t", "int", "_DWORD", "UINT_PTR", "DWORD",
669 "WPARAM", "LPARAM", "UINT", "__int32",
670 "LONG", "HIMC", "BOOL", "size_t",
671 "float",
672 };
673 static const char *word_types[] = {
674 "uint16_t", "int16_t", "_WORD", "WORD",
675 "unsigned __int16", "__int16",
676 };
677 static const char *byte_types[] = {
678 "uint8_t", "int8_t", "char",
679 "unsigned __int8", "__int8", "BYTE", "_BYTE",
680 "CHAR", "_UNKNOWN",
681 // structures.. deal the same as with _UNKNOWN for now
682 "CRITICAL_SECTION",
683 };
684 const char *n;
685 int i;
686
687 if (c_type->is_ptr) {
688 *lmod = OPLM_DWORD;
689 return 1;
690 }
691
692 n = skip_type_mod(c_type->name);
693
694 for (i = 0; i < ARRAY_SIZE(dword_types); i++) {
695 if (IS(n, dword_types[i])) {
696 *lmod = OPLM_DWORD;
697 return 1;
698 }
699 }
700
701 for (i = 0; i < ARRAY_SIZE(word_types); i++) {
702 if (IS(n, word_types[i])) {
703 *lmod = OPLM_WORD;
704 return 1;
705 }
706 }
707
708 for (i = 0; i < ARRAY_SIZE(byte_types); i++) {
709 if (IS(n, byte_types[i])) {
710 *lmod = OPLM_BYTE;
711 return 1;
712 }
713 }
714
715 for (i = 0; i < ARRAY_SIZE(qword_types); i++) {
716 if (IS(n, qword_types[i])) {
717 *lmod = OPLM_QWORD;
718 return 1;
719 }
720 }
721
722 return 0;
723}
724
725static char *default_cast_to(char *buf, size_t buf_size,
726 struct parsed_opr *opr)
727{
728 buf[0] = 0;
729
730 if (!opr->is_ptr || strchr(opr->name, '['))
731 return buf;
732 if (opr->pp == NULL || opr->pp->type.name == NULL
733 || opr->pp->is_fptr)
734 {
735 snprintf(buf, buf_size, "%s", "(void *)");
736 return buf;
737 }
738
739 snprintf(buf, buf_size, "(%s)", opr->pp->type.name);
740 return buf;
741}
742
743static enum opr_type lmod_from_directive(const char *d)
744{
745 if (IS(d, "dd"))
746 return OPLM_DWORD;
747 else if (IS(d, "dw"))
748 return OPLM_WORD;
749 else if (IS(d, "db"))
750 return OPLM_BYTE;
751
752 aerr("unhandled directive: '%s'\n", d);
753 return OPLM_UNSPEC;
754}
755
756static void setup_reg_opr(struct parsed_opr *opr, int reg, enum opr_lenmod lmod,
757 int *regmask)
758{
759 opr->type = OPT_REG;
760 opr->reg = reg;
761 opr->lmod = lmod;
762 *regmask |= 1 << reg;
763}
764
765static struct parsed_equ *equ_find(struct parsed_op *po, const char *name,
766 int *extra_offs);
767
768static int parse_operand(struct parsed_opr *opr,
769 int *regmask, int *regmask_indirect,
770 char words[16][256], int wordc, int w, unsigned int op_flags)
771{
772 const struct parsed_proto *pp = NULL;
773 enum opr_lenmod tmplmod;
774 unsigned long number;
775 char buf[256];
776 int ret, len;
777 int wordc_in;
778 char *p;
779 int i;
780
781 if (w >= wordc)
782 aerr("parse_operand w %d, wordc %d\n", w, wordc);
783
784 opr->reg = xUNSPEC;
785
786 for (i = w; i < wordc; i++) {
787 len = strlen(words[i]);
788 if (words[i][len - 1] == ',') {
789 words[i][len - 1] = 0;
790 wordc = i + 1;
791 break;
792 }
793 }
794
795 wordc_in = wordc - w;
796
797 if ((op_flags & OPF_JMP) && wordc_in > 0
798 && !('0' <= words[w][0] && words[w][0] <= '9'))
799 {
800 const char *label = NULL;
801
802 if (wordc_in == 3 && !strncmp(words[w], "near", 4)
803 && IS(words[w + 1], "ptr"))
804 label = words[w + 2];
805 else if (wordc_in == 2 && IS(words[w], "short"))
806 label = words[w + 1];
807 else if (wordc_in == 1
808 && strchr(words[w], '[') == NULL
809 && parse_reg(&tmplmod, words[w]) < 0)
810 label = words[w];
811
812 if (label != NULL) {
813 opr->type = OPT_LABEL;
814 ret = check_segment_prefix(label);
815 if (ret != 0) {
816 opr->segment = ret;
817 label += 3;
818 }
819 strcpy(opr->name, label);
820 return wordc;
821 }
822 }
823
824 if (wordc_in >= 3) {
825 if (IS(words[w + 1], "ptr")) {
826 if (IS(words[w], "dword"))
827 opr->lmod = OPLM_DWORD;
828 else if (IS(words[w], "word"))
829 opr->lmod = OPLM_WORD;
830 else if (IS(words[w], "byte"))
831 opr->lmod = OPLM_BYTE;
832 else if (IS(words[w], "qword"))
833 opr->lmod = OPLM_QWORD;
834 else
835 aerr("type parsing failed\n");
836 w += 2;
837 wordc_in = wordc - w;
838 }
839 }
840
841 if (wordc_in == 2) {
842 if (IS(words[w], "offset")) {
843 opr->type = OPT_OFFSET;
844 opr->lmod = OPLM_DWORD;
845 strcpy(opr->name, words[w + 1]);
846 pp = proto_parse(g_fhdr, opr->name, 1);
847 goto do_label;
848 }
849 if (IS(words[w], "(offset")) {
850 p = strchr(words[w + 1], ')');
851 if (p == NULL)
852 aerr("parse of bracketed offset failed\n");
853 *p = 0;
854 opr->type = OPT_OFFSET;
855 strcpy(opr->name, words[w + 1]);
856 return wordc;
857 }
858 }
859
860 if (wordc_in != 1)
861 aerr("parse_operand 1 word expected\n");
862
863 ret = check_segment_prefix(words[w]);
864 if (ret != 0) {
865 opr->segment = ret;
866 memmove(words[w], words[w] + 3, strlen(words[w]) - 2);
867 if (ret == SEG_FS && IS(words[w], "0"))
868 g_seh_found = 1;
869 }
870 strcpy(opr->name, words[w]);
871
872 if (words[w][0] == '[') {
873 opr->type = OPT_REGMEM;
874 ret = sscanf(words[w], "[%[^]]]", opr->name);
875 if (ret != 1)
876 aerr("[] parse failure\n");
877
878 parse_indmode(opr->name, regmask_indirect, 1);
879 if (opr->lmod == OPLM_UNSPEC
880 && parse_stack_el(opr->name, NULL, NULL, 1))
881 {
882 // might be an equ
883 struct parsed_equ *eq =
884 equ_find(NULL, parse_stack_el(opr->name, NULL, NULL, 1), &i);
885 if (eq)
886 opr->lmod = eq->lmod;
887
888 // might be unaligned access
889 g_func_lmods |= 1 << OPLM_BYTE;
890 }
891 return wordc;
892 }
893 else if (strchr(words[w], '[')) {
894 // label[reg] form
895 p = strchr(words[w], '[');
896 opr->type = OPT_REGMEM;
897 parse_indmode(p, regmask_indirect, 0);
898 strncpy(buf, words[w], p - words[w]);
899 buf[p - words[w]] = 0;
900 pp = proto_parse(g_fhdr, buf, 1);
901 goto do_label;
902 }
903 else if (('0' <= words[w][0] && words[w][0] <= '9')
904 || words[w][0] == '-')
905 {
906 number = parse_number(words[w], 0);
907 opr->type = OPT_CONST;
908 opr->val = number;
909 printf_number(opr->name, sizeof(opr->name), number);
910 return wordc;
911 }
912
913 ret = parse_reg(&tmplmod, opr->name);
914 if (ret >= 0) {
915 setup_reg_opr(opr, ret, tmplmod, regmask);
916 return wordc;
917 }
918
919 // most likely var in data segment
920 opr->type = OPT_LABEL;
921 pp = proto_parse(g_fhdr, opr->name, g_quiet_pp);
922
923do_label:
924 if (pp != NULL) {
925 if (pp->is_fptr || pp->is_func) {
926 opr->lmod = OPLM_DWORD;
927 opr->is_ptr = 1;
928 }
929 else {
930 tmplmod = OPLM_UNSPEC;
931 if (!guess_lmod_from_c_type(&tmplmod, &pp->type))
932 anote("unhandled C type '%s' for '%s'\n",
933 pp->type.name, opr->name);
934
935 if (opr->lmod == OPLM_UNSPEC) {
936 opr->lmod = tmplmod;
937 opr->type_from_var = 1;
938 }
939 else if (opr->lmod != tmplmod) {
940 opr->size_mismatch = 1;
941 if (tmplmod < opr->lmod)
942 opr->size_lt = 1;
943 }
944 opr->is_ptr = pp->type.is_ptr;
945 }
946 opr->is_array = pp->type.is_array;
947 }
948 opr->pp = pp;
949
950 if (opr->lmod == OPLM_UNSPEC)
951 guess_lmod_from_name(opr);
952 return wordc;
953}
954
955static const struct {
956 const char *name;
957 unsigned int flags;
958} pref_table[] = {
959 { "rep", OPF_REP },
960 { "repe", OPF_REP|OPF_REPZ },
961 { "repz", OPF_REP|OPF_REPZ },
962 { "repne", OPF_REP|OPF_REPNZ },
963 { "repnz", OPF_REP|OPF_REPNZ },
964 { "lock", OPF_LOCK },
965};
966
967#define OPF_CJMP_CC (OPF_JMP|OPF_CJMP|OPF_CC)
968
969static const struct {
970 const char *name;
971 enum op_op op;
972 unsigned short minopr;
973 unsigned short maxopr;
974 unsigned int flags;
975 unsigned char pfo;
976 unsigned char pfo_inv;
977} op_table[] = {
978 { "nop", OP_NOP, 0, 0, 0 },
979 { "push", OP_PUSH, 1, 1, 0 },
980 { "pop", OP_POP, 1, 1, OPF_DATA },
981 { "pusha",OP_PUSHA, 0, 0, 0 },
982 { "popa", OP_POPA, 0, 0, OPF_DATA },
983 { "leave",OP_LEAVE, 0, 0, OPF_DATA },
984 { "mov" , OP_MOV, 2, 2, OPF_DATA },
985 { "lea", OP_LEA, 2, 2, OPF_DATA },
986 { "movzx",OP_MOVZX, 2, 2, OPF_DATA },
987 { "movsx",OP_MOVSX, 2, 2, OPF_DATA },
988 { "xchg", OP_XCHG, 2, 2, OPF_DATA },
989 { "not", OP_NOT, 1, 1, OPF_DATA },
990 { "xlat", OP_XLAT, 0, 0, OPF_DATA },
991 { "cdq", OP_CDQ, 0, 0, OPF_DATA },
992 { "bswap",OP_BSWAP, 1, 1, OPF_DATA },
993 { "lodsb",OP_LODS, 0, 0, OPF_DATA },
994 { "lodsw",OP_LODS, 0, 0, OPF_DATA },
995 { "lodsd",OP_LODS, 0, 0, OPF_DATA },
996 { "stosb",OP_STOS, 0, 0, OPF_DATA },
997 { "stosw",OP_STOS, 0, 0, OPF_DATA },
998 { "stosd",OP_STOS, 0, 0, OPF_DATA },
999 { "movsb",OP_MOVS, 0, 0, OPF_DATA },
1000 { "movsw",OP_MOVS, 0, 0, OPF_DATA },
1001 { "movsd",OP_MOVS, 0, 0, OPF_DATA },
1002 { "cmpsb",OP_CMPS, 0, 0, OPF_DATA|OPF_FLAGS },
1003 { "cmpsw",OP_CMPS, 0, 0, OPF_DATA|OPF_FLAGS },
1004 { "cmpsd",OP_CMPS, 0, 0, OPF_DATA|OPF_FLAGS },
1005 { "scasb",OP_SCAS, 0, 0, OPF_DATA|OPF_FLAGS },
1006 { "scasw",OP_SCAS, 0, 0, OPF_DATA|OPF_FLAGS },
1007 { "scasd",OP_SCAS, 0, 0, OPF_DATA|OPF_FLAGS },
1008 { "rdtsc",OP_RDTSC, 0, 0, OPF_DATA },
1009 { "cpuid",OP_CPUID, 0, 0, OPF_DATA },
1010 { "std", OP_STD, 0, 0, OPF_DATA }, // special flag
1011 { "cld", OP_CLD, 0, 0, OPF_DATA },
1012 { "add", OP_ADD, 2, 2, OPF_DATA|OPF_FLAGS },
1013 { "sub", OP_SUB, 2, 2, OPF_DATA|OPF_FLAGS },
1014 { "and", OP_AND, 2, 2, OPF_DATA|OPF_FLAGS },
1015 { "or", OP_OR, 2, 2, OPF_DATA|OPF_FLAGS },
1016 { "xor", OP_XOR, 2, 2, OPF_DATA|OPF_FLAGS },
1017 { "shl", OP_SHL, 2, 2, OPF_DATA|OPF_FLAGS },
1018 { "shr", OP_SHR, 2, 2, OPF_DATA|OPF_FLAGS },
1019 { "sal", OP_SHL, 2, 2, OPF_DATA|OPF_FLAGS },
1020 { "sar", OP_SAR, 2, 2, OPF_DATA|OPF_FLAGS },
1021 { "shld", OP_SHLD, 3, 3, OPF_DATA|OPF_FLAGS },
1022 { "shrd", OP_SHRD, 3, 3, OPF_DATA|OPF_FLAGS },
1023 { "rol", OP_ROL, 2, 2, OPF_DATA|OPF_FLAGS },
1024 { "ror", OP_ROR, 2, 2, OPF_DATA|OPF_FLAGS },
1025 { "rcl", OP_RCL, 2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
1026 { "rcr", OP_RCR, 2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
1027 { "adc", OP_ADC, 2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
1028 { "sbb", OP_SBB, 2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
1029 { "bsf", OP_BSF, 2, 2, OPF_DATA|OPF_FLAGS },
1030 { "bsr", OP_BSR, 2, 2, OPF_DATA|OPF_FLAGS },
1031 { "inc", OP_INC, 1, 1, OPF_DATA|OPF_FLAGS },
1032 { "dec", OP_DEC, 1, 1, OPF_DATA|OPF_FLAGS },
1033 { "neg", OP_NEG, 1, 1, OPF_DATA|OPF_FLAGS },
1034 { "mul", OP_MUL, 1, 1, OPF_DATA|OPF_FLAGS },
1035 { "imul", OP_IMUL, 1, 3, OPF_DATA|OPF_FLAGS },
1036 { "div", OP_DIV, 1, 1, OPF_DATA|OPF_FLAGS },
1037 { "idiv", OP_IDIV, 1, 1, OPF_DATA|OPF_FLAGS },
1038 { "test", OP_TEST, 2, 2, OPF_FLAGS },
1039 { "cmp", OP_CMP, 2, 2, OPF_FLAGS },
1040 { "retn", OP_RET, 0, 1, OPF_TAIL },
1041 { "call", OP_CALL, 1, 1, OPF_JMP|OPF_DATA|OPF_FLAGS },
1042 { "jmp", OP_JMP, 1, 1, OPF_JMP },
1043 { "jecxz",OP_JECXZ, 1, 1, OPF_JMP|OPF_CJMP },
1044 { "loop", OP_LOOP, 1, 1, OPF_JMP|OPF_CJMP|OPF_DATA },
1045 { "jo", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_O, 0 }, // 70 OF=1
1046 { "jno", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_O, 1 }, // 71 OF=0
1047 { "jc", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_C, 0 }, // 72 CF=1
1048 { "jb", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_C, 0 }, // 72
1049 { "jnc", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_C, 1 }, // 73 CF=0
1050 { "jnb", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_C, 1 }, // 73
1051 { "jae", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_C, 1 }, // 73
1052 { "jz", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_Z, 0 }, // 74 ZF=1
1053 { "je", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_Z, 0 }, // 74
1054 { "jnz", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_Z, 1 }, // 75 ZF=0
1055 { "jne", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_Z, 1 }, // 75
1056 { "jbe", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_BE, 0 }, // 76 CF=1||ZF=1
1057 { "jna", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_BE, 0 }, // 76
1058 { "ja", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_BE, 1 }, // 77 CF=0&&ZF=0
1059 { "jnbe", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_BE, 1 }, // 77
1060 { "js", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_S, 0 }, // 78 SF=1
1061 { "jns", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_S, 1 }, // 79 SF=0
1062 { "jp", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_P, 0 }, // 7a PF=1
1063 { "jpe", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_P, 0 }, // 7a
1064 { "jnp", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_P, 1 }, // 7b PF=0
1065 { "jpo", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_P, 1 }, // 7b
1066 { "jl", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_L, 0 }, // 7c SF!=OF
1067 { "jnge", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_L, 0 }, // 7c
1068 { "jge", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_L, 1 }, // 7d SF=OF
1069 { "jnl", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_L, 1 }, // 7d
1070 { "jle", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_LE, 0 }, // 7e ZF=1||SF!=OF
1071 { "jng", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_LE, 0 }, // 7e
1072 { "jg", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_LE, 1 }, // 7f ZF=0&&SF=OF
1073 { "jnle", OP_JCC, 1, 1, OPF_CJMP_CC, PFO_LE, 1 }, // 7f
1074 { "seto", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_O, 0 },
1075 { "setno", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_O, 1 },
1076 { "setc", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_C, 0 },
1077 { "setb", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_C, 0 },
1078 { "setnc", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_C, 1 },
1079 { "setae", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_C, 1 },
1080 { "setnb", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_C, 1 },
1081 { "setz", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_Z, 0 },
1082 { "sete", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_Z, 0 },
1083 { "setnz", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_Z, 1 },
1084 { "setne", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_Z, 1 },
1085 { "setbe", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_BE, 0 },
1086 { "setna", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_BE, 0 },
1087 { "seta", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_BE, 1 },
1088 { "setnbe", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_BE, 1 },
1089 { "sets", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_S, 0 },
1090 { "setns", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_S, 1 },
1091 { "setp", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_P, 0 },
1092 { "setpe", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_P, 0 },
1093 { "setnp", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_P, 1 },
1094 { "setpo", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_P, 1 },
1095 { "setl", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_L, 0 },
1096 { "setnge", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_L, 0 },
1097 { "setge", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_L, 1 },
1098 { "setnl", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_L, 1 },
1099 { "setle", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_LE, 0 },
1100 { "setng", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_LE, 0 },
1101 { "setg", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_LE, 1 },
1102 { "setnle", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_LE, 1 },
1103 // x87
1104 { "fld", OP_FLD, 1, 1, OPF_FPUSH },
1105 { "fild", OP_FILD, 1, 1, OPF_FPUSH|OPF_FINT },
1106 { "fld1", OP_FLDc, 0, 0, OPF_FPUSH },
1107 { "fldl2t", OP_FLDc, 0, 0, OPF_FPUSH },
1108 { "fldl2e", OP_FLDc, 0, 0, OPF_FPUSH },
1109 { "fldpi", OP_FLDc, 0, 0, OPF_FPUSH },
1110 { "fldlg2", OP_FLDc, 0, 0, OPF_FPUSH },
1111 { "fldln2", OP_FLDc, 0, 0, OPF_FPUSH },
1112 { "fldz", OP_FLDc, 0, 0, OPF_FPUSH },
1113 { "fst", OP_FST, 1, 1, 0 },
1114 { "fstp", OP_FST, 1, 1, OPF_FPOP },
1115 { "fist", OP_FIST, 1, 1, OPF_FINT },
1116 { "fistp", OP_FIST, 1, 1, OPF_FPOP|OPF_FINT },
1117 { "fabs", OP_FABS, 0, 0, 0 },
1118 { "fadd", OP_FADD, 0, 2, 0 },
1119 { "faddp", OP_FADD, 0, 2, OPF_FPOP },
1120 { "fdiv", OP_FDIV, 0, 2, 0 },
1121 { "fdivp", OP_FDIV, 0, 2, OPF_FPOP },
1122 { "fmul", OP_FMUL, 0, 2, 0 },
1123 { "fmulp", OP_FMUL, 0, 2, OPF_FPOP },
1124 { "fsub", OP_FSUB, 0, 2, 0 },
1125 { "fsubp", OP_FSUB, 0, 2, OPF_FPOP },
1126 { "fdivr", OP_FDIVR, 0, 2, 0 },
1127 { "fdivrp", OP_FDIVR, 0, 2, OPF_FPOP },
1128 { "fsubr", OP_FSUBR, 0, 2, 0 },
1129 { "fsubrp", OP_FSUBR, 0, 2, OPF_FPOP },
1130 { "fiadd", OP_FIADD, 1, 1, OPF_FINT },
1131 { "fidiv", OP_FIDIV, 1, 1, OPF_FINT },
1132 { "fimul", OP_FIMUL, 1, 1, OPF_FINT },
1133 { "fisub", OP_FISUB, 1, 1, OPF_FINT },
1134 { "fidivr", OP_FIDIVR, 1, 1, OPF_FINT },
1135 { "fisubr", OP_FISUBR, 1, 1, OPF_FINT },
1136 { "fcom", OP_FCOM, 0, 1, 0 },
1137 { "fcomp", OP_FCOM, 0, 1, OPF_FPOP },
1138 { "fcompp", OP_FCOM, 0, 0, OPF_FPOPP },
1139 { "fucom", OP_FCOM, 0, 1, 0 },
1140 { "fucomp", OP_FCOM, 0, 1, OPF_FPOP },
1141 { "fucompp",OP_FCOM, 0, 0, OPF_FPOPP },
1142 { "fnstsw", OP_FNSTSW, 1, 1, OPF_DATA },
1143 { "fchs", OP_FCHS, 0, 0, 0 },
1144 { "fcos", OP_FCOS, 0, 0, 0 },
1145 { "fpatan", OP_FPATAN, 0, 0, OPF_FPOP },
1146 { "fptan", OP_FPTAN, 0, 0, OPF_FPUSH },
1147 { "fsin", OP_FSIN, 0, 0, 0 },
1148 { "fsqrt", OP_FSQRT, 0, 0, 0 },
1149 { "fxch", OP_FXCH, 1, 1, 0 },
1150 { "fyl2x", OP_FYL2X, 0, 0, OPF_FPOP },
1151 // mmx
1152 { "emms", OP_EMMS, 0, 0, OPF_DATA },
1153 { "movq", OP_MOV, 2, 2, OPF_DATA },
1154 // pseudo-ops for lib calls
1155 { "_allshl",OPP_ALLSHL },
1156 { "_allshr",OPP_ALLSHR },
1157 { "_ftol", OPP_FTOL },
1158 { "_CIpow", OPP_CIPOW },
1159 { "abort", OPP_ABORT },
1160 // must be last
1161 { "ud2", OP_UD2 },
1162};
1163
1164static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
1165{
1166 enum opr_lenmod lmod = OPLM_UNSPEC;
1167 int prefix_flags = 0;
1168 int regmask_ind;
1169 int regmask;
1170 int op_w = 0;
1171 int opr = 0;
1172 int w = 0;
1173 int i, j;
1174
1175 for (i = 0; i < ARRAY_SIZE(pref_table); i++) {
1176 if (IS(words[w], pref_table[i].name)) {
1177 prefix_flags = pref_table[i].flags;
1178 break;
1179 }
1180 }
1181
1182 if (prefix_flags) {
1183 if (wordc <= 1)
1184 aerr("lone prefix: '%s'\n", words[0]);
1185 w++;
1186 }
1187
1188 op_w = w;
1189 for (i = 0; i < ARRAY_SIZE(op_table); i++) {
1190 if (IS(words[w], op_table[i].name))
1191 break;
1192 }
1193
1194 if (i == ARRAY_SIZE(op_table)) {
1195 if (!g_skip_func)
1196 aerr("unhandled op: '%s'\n", words[0]);
1197 i--; // OP_UD2
1198 }
1199 w++;
1200
1201 op->op = op_table[i].op;
1202 op->flags = op_table[i].flags | prefix_flags;
1203 op->pfo = op_table[i].pfo;
1204 op->pfo_inv = op_table[i].pfo_inv;
1205 op->regmask_src = op->regmask_dst = 0;
1206 op->asmln = asmln;
1207
1208 if (op->op == OP_UD2)
1209 return;
1210
1211 for (opr = 0; opr < op_table[i].maxopr; opr++) {
1212 if (opr >= op_table[i].minopr && w >= wordc)
1213 break;
1214
1215 regmask = regmask_ind = 0;
1216 w = parse_operand(&op->operand[opr], &regmask, &regmask_ind,
1217 words, wordc, w, op->flags);
1218
1219 if (opr == 0 && (op->flags & OPF_DATA))
1220 op->regmask_dst = regmask;
1221 else
1222 op->regmask_src |= regmask;
1223 op->regmask_src |= regmask_ind;
1224
1225 if (op->operand[opr].lmod != OPLM_UNSPEC)
1226 g_func_lmods |= 1 << op->operand[opr].lmod;
1227 }
1228
1229 if (w < wordc)
1230 aerr("parse_op %s incomplete: %d/%d\n",
1231 words[0], w, wordc);
1232
1233 // special cases
1234 op->operand_cnt = opr;
1235 if (!strncmp(op_table[i].name, "set", 3))
1236 op->operand[0].lmod = OPLM_BYTE;
1237
1238 switch (op->op) {
1239 // first operand is not dst
1240 case OP_CMP:
1241 case OP_TEST:
1242 op->regmask_src |= op->regmask_dst;
1243 op->regmask_dst = 0;
1244 break;
1245
1246 // first operand is src too
1247 case OP_NOT:
1248 case OP_ADD:
1249 case OP_AND:
1250 case OP_OR:
1251 case OP_RCL:
1252 case OP_RCR:
1253 case OP_ADC:
1254 case OP_INC:
1255 case OP_DEC:
1256 case OP_NEG:
1257 case OP_BSWAP:
1258 // more below..
1259 op->regmask_src |= op->regmask_dst;
1260 break;
1261
1262 // special
1263 case OP_XCHG:
1264 op->regmask_src |= op->regmask_dst;
1265 op->regmask_dst |= op->regmask_src;
1266 goto check_align;
1267
1268 case OP_SUB:
1269 case OP_SBB:
1270 case OP_XOR:
1271 if (op->operand[0].type == OPT_REG && op->operand[1].type == OPT_REG
1272 && op->operand[0].lmod == op->operand[1].lmod
1273 && op->operand[0].reg == op->operand[1].reg
1274 && IS(op->operand[0].name, op->operand[1].name)) // ! ah, al..
1275 {
1276 op->regmask_src = 0;
1277 }
1278 else
1279 op->regmask_src |= op->regmask_dst;
1280 break;
1281
1282 // ops with implicit argumets
1283 case OP_XLAT:
1284 op->operand_cnt = 2;
1285 setup_reg_opr(&op->operand[0], xAX, OPLM_BYTE, &op->regmask_src);
1286 op->regmask_dst = op->regmask_src;
1287 setup_reg_opr(&op->operand[1], xBX, OPLM_DWORD, &op->regmask_src);
1288 break;
1289
1290 case OP_CDQ:
1291 op->operand_cnt = 2;
1292 setup_reg_opr(&op->operand[0], xDX, OPLM_DWORD, &op->regmask_dst);
1293 setup_reg_opr(&op->operand[1], xAX, OPLM_DWORD, &op->regmask_src);
1294 break;
1295
1296 case OP_LODS:
1297 case OP_STOS:
1298 case OP_SCAS:
1299 if (words[op_w][4] == 'b')
1300 lmod = OPLM_BYTE;
1301 else if (words[op_w][4] == 'w')
1302 lmod = OPLM_WORD;
1303 else if (words[op_w][4] == 'd')
1304 lmod = OPLM_DWORD;
1305 j = 0;
1306 op->regmask_src = 0;
1307 setup_reg_opr(&op->operand[j++], op->op == OP_LODS ? xSI : xDI,
1308 OPLM_DWORD, &op->regmask_src);
1309 op->regmask_dst = op->regmask_src;
1310 setup_reg_opr(&op->operand[j++], xAX, lmod,
1311 op->op == OP_LODS ? &op->regmask_dst : &op->regmask_src);
1312 if (op->flags & OPF_REP) {
1313 setup_reg_opr(&op->operand[j++], xCX, OPLM_DWORD, &op->regmask_src);
1314 op->regmask_dst |= 1 << xCX;
1315 }
1316 op->operand_cnt = j;
1317 break;
1318
1319 case OP_MOVS:
1320 case OP_CMPS:
1321 if (words[op_w][4] == 'b')
1322 lmod = OPLM_BYTE;
1323 else if (words[op_w][4] == 'w')
1324 lmod = OPLM_WORD;
1325 else if (words[op_w][4] == 'd')
1326 lmod = OPLM_DWORD;
1327 j = 0;
1328 op->regmask_src = 0;
1329 // note: lmod is not correct, don't have where to place it
1330 setup_reg_opr(&op->operand[j++], xDI, lmod, &op->regmask_src);
1331 setup_reg_opr(&op->operand[j++], xSI, OPLM_DWORD, &op->regmask_src);
1332 if (op->flags & OPF_REP)
1333 setup_reg_opr(&op->operand[j++], xCX, OPLM_DWORD, &op->regmask_src);
1334 op->operand_cnt = j;
1335 op->regmask_dst = op->regmask_src;
1336 break;
1337
1338 case OP_RDTSC:
1339 op->regmask_dst = mxAX | mxDX;
1340 break;
1341
1342 case OP_CPUID:
1343 // for now, ignore ecx dep for eax={4,7,b,d}
1344 op->regmask_src = mxAX;
1345 op->regmask_dst = mxAX | mxBX | mxCX | mxDX;
1346 break;
1347
1348 case OP_LOOP:
1349 op->regmask_dst = 1 << xCX;
1350 // fallthrough
1351 case OP_JECXZ:
1352 op->operand_cnt = 2;
1353 op->regmask_src = 1 << xCX;
1354 op->operand[1].type = OPT_REG;
1355 op->operand[1].reg = xCX;
1356 op->operand[1].lmod = OPLM_DWORD;
1357 break;
1358
1359 case OP_IMUL:
1360 if (op->operand_cnt == 2) {
1361 if (op->operand[0].type != OPT_REG)
1362 aerr("reg expected\n");
1363 op->regmask_src |= 1 << op->operand[0].reg;
1364 }
1365 if (op->operand_cnt != 1)
1366 break;
1367 // fallthrough
1368 case OP_MUL:
1369 // singleop mul
1370 if (op->operand[0].lmod == OPLM_UNSPEC)
1371 op->operand[0].lmod = OPLM_DWORD;
1372 op->regmask_src = mxAX | op->regmask_dst;
1373 op->regmask_dst = mxAX;
1374 if (op->operand[0].lmod != OPLM_BYTE)
1375 op->regmask_dst |= mxDX;
1376 break;
1377
1378 case OP_DIV:
1379 case OP_IDIV:
1380 // we could set up operands for edx:eax, but there is no real need to
1381 // (see is_opr_modified())
1382 if (op->operand[0].lmod == OPLM_UNSPEC)
1383 op->operand[0].lmod = OPLM_DWORD;
1384 op->regmask_src = mxAX | op->regmask_dst;
1385 op->regmask_dst = mxAX;
1386 if (op->operand[0].lmod != OPLM_BYTE) {
1387 op->regmask_src |= mxDX;
1388 op->regmask_dst |= mxDX;
1389 }
1390 break;
1391
1392 case OP_SHL:
1393 case OP_SHR:
1394 case OP_SAR:
1395 case OP_ROL:
1396 case OP_ROR:
1397 op->regmask_src |= op->regmask_dst;
1398 if (op->operand[1].lmod == OPLM_UNSPEC)
1399 op->operand[1].lmod = OPLM_BYTE;
1400 break;
1401
1402 case OP_SHLD:
1403 case OP_SHRD:
1404 op->regmask_src |= op->regmask_dst;
1405 if (op->operand[2].lmod == OPLM_UNSPEC)
1406 op->operand[2].lmod = OPLM_BYTE;
1407 break;
1408
1409 case OP_PUSH:
1410 op->regmask_src |= op->regmask_dst;
1411 op->regmask_dst = 0;
1412 if (op->operand[0].lmod == OPLM_UNSPEC
1413 && (op->operand[0].type == OPT_CONST
1414 || op->operand[0].type == OPT_OFFSET
1415 || op->operand[0].type == OPT_LABEL))
1416 op->operand[0].lmod = OPLM_DWORD;
1417 break;
1418
1419 // alignment
1420 case OP_MOV:
1421 check_align:
1422 if (op->operand[0].type == OPT_REG && op->operand[1].type == OPT_REG
1423 && op->operand[0].lmod == op->operand[1].lmod
1424 && op->operand[0].reg == op->operand[1].reg
1425 && IS(op->operand[0].name, op->operand[1].name)) // ! ah, al..
1426 {
1427 op->flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
1428 op->regmask_src = op->regmask_dst = 0;
1429 }
1430 break;
1431
1432 case OP_LEA:
1433 if (op->operand[0].type == OPT_REG
1434 && op->operand[1].type == OPT_REGMEM)
1435 {
1436 char buf[16];
1437 snprintf(buf, sizeof(buf), "%s+0", op->operand[0].name);
1438 if (IS(buf, op->operand[1].name))
1439 op->flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
1440 }
1441 break;
1442
1443 case OP_CALL:
1444 // needed because of OPF_DATA
1445 op->regmask_src |= op->regmask_dst;
1446 // trashed regs must be explicitly detected later
1447 op->regmask_dst = 0;
1448 break;
1449
1450 case OP_LEAVE:
1451 op->regmask_dst = (1 << xBP) | (1 << xSP);
1452 op->regmask_src = 1 << xBP;
1453 break;
1454
1455 case OP_FLD:
1456 case OP_FILD:
1457 op->regmask_dst |= mxST0;
1458 break;
1459
1460 case OP_FLDc:
1461 op->regmask_dst |= mxST0;
1462 if (IS(words[op_w] + 3, "1"))
1463 op->operand[0].val = X87_CONST_1;
1464 else if (IS(words[op_w] + 3, "l2t"))
1465 op->operand[0].val = X87_CONST_L2T;
1466 else if (IS(words[op_w] + 3, "l2e"))
1467 op->operand[0].val = X87_CONST_L2E;
1468 else if (IS(words[op_w] + 3, "pi"))
1469 op->operand[0].val = X87_CONST_PI;
1470 else if (IS(words[op_w] + 3, "lg2"))
1471 op->operand[0].val = X87_CONST_LG2;
1472 else if (IS(words[op_w] + 3, "ln2"))
1473 op->operand[0].val = X87_CONST_LN2;
1474 else if (IS(words[op_w] + 3, "z"))
1475 op->operand[0].val = X87_CONST_Z;
1476 else
1477 aerr("fld what?\n");
1478 break;
1479
1480 case OP_FST:
1481 case OP_FIST:
1482 op->regmask_src |= mxST0;
1483 break;
1484
1485 case OP_FADD:
1486 case OP_FDIV:
1487 case OP_FMUL:
1488 case OP_FSUB:
1489 case OP_FDIVR:
1490 case OP_FSUBR:
1491 op->regmask_src |= mxST0;
1492 if (op->operand_cnt == 2)
1493 op->regmask_src |= op->regmask_dst;
1494 else if (op->operand_cnt == 1) {
1495 memcpy(&op->operand[1], &op->operand[0], sizeof(op->operand[1]));
1496 op->operand[0].type = OPT_REG;
1497 op->operand[0].lmod = OPLM_QWORD;
1498 op->operand[0].reg = xST0;
1499 op->regmask_dst |= mxST0;
1500 }
1501 else
1502 // IDA doesn't use this
1503 aerr("no operands?\n");
1504 break;
1505
1506 case OP_FIADD:
1507 case OP_FIDIV:
1508 case OP_FIMUL:
1509 case OP_FISUB:
1510 case OP_FIDIVR:
1511 case OP_FISUBR:
1512 case OP_FABS:
1513 case OP_FCHS:
1514 case OP_FCOS:
1515 case OP_FSIN:
1516 case OP_FSQRT:
1517 case OP_FXCH:
1518 op->regmask_src |= mxST0;
1519 op->regmask_dst |= mxST0;
1520 break;
1521
1522 case OP_FPATAN:
1523 case OP_FYL2X:
1524 op->regmask_src |= mxST0 | mxST1;
1525 op->regmask_dst |= mxST0;
1526 break;
1527
1528 case OP_FPTAN:
1529 aerr("TODO\n");
1530 break;
1531
1532 case OP_FCOM:
1533 op->regmask_src |= mxST0;
1534 if (op->operand_cnt == 0) {
1535 op->operand_cnt = 1;
1536 op->operand[0].type = OPT_REG;
1537 op->operand[0].lmod = OPLM_QWORD;
1538 op->operand[0].reg = xST1;
1539 op->regmask_src |= mxST1;
1540 }
1541 break;
1542
1543 default:
1544 break;
1545 }
1546
1547 if (op->operand[0].type == OPT_REG
1548 && op->operand[1].type == OPT_CONST)
1549 {
1550 struct parsed_opr *op1 = &op->operand[1];
1551 if ((op->op == OP_AND && op1->val == 0)
1552 || (op->op == OP_OR
1553 && (op1->val == ~0
1554 || (op->operand[0].lmod == OPLM_WORD && op1->val == 0xffff)
1555 || (op->operand[0].lmod == OPLM_BYTE && op1->val == 0xff))))
1556 {
1557 op->regmask_src = 0;
1558 }
1559 }
1560}
1561
1562static const char *op_name(struct parsed_op *po)
1563{
1564 static char buf[16];
1565 char *p;
1566 int i;
1567
1568 if (po->op == OP_JCC || po->op == OP_SCC) {
1569 p = buf;
1570 *p++ = (po->op == OP_JCC) ? 'j' : 's';
1571 if (po->pfo_inv)
1572 *p++ = 'n';
1573 strcpy(p, parsed_flag_op_names[po->pfo]);
1574 return buf;
1575 }
1576
1577 for (i = 0; i < ARRAY_SIZE(op_table); i++)
1578 if (op_table[i].op == po->op)
1579 return op_table[i].name;
1580
1581 return "???";
1582}
1583
1584// debug
1585static const char *dump_op(struct parsed_op *po)
1586{
1587 static char out[128];
1588 char *p = out;
1589 int i;
1590
1591 if (po == NULL)
1592 return "???";
1593
1594 snprintf(out, sizeof(out), "%s", op_name(po));
1595 for (i = 0; i < po->operand_cnt; i++) {
1596 p += strlen(p);
1597 if (i > 0)
1598 *p++ = ',';
1599 snprintf(p, sizeof(out) - (p - out),
1600 po->operand[i].type == OPT_REGMEM ? " [%s]" : " %s",
1601 po->operand[i].name);
1602 }
1603
1604 return out;
1605}
1606
1607static const char *lmod_type_u(struct parsed_op *po,
1608 enum opr_lenmod lmod)
1609{
1610 switch (lmod) {
1611 case OPLM_QWORD:
1612 return "u64";
1613 case OPLM_DWORD:
1614 return "u32";
1615 case OPLM_WORD:
1616 return "u16";
1617 case OPLM_BYTE:
1618 return "u8";
1619 default:
1620 ferr(po, "invalid lmod: %d\n", lmod);
1621 return "(_invalid_)";
1622 }
1623}
1624
1625static const char *lmod_cast_u(struct parsed_op *po,
1626 enum opr_lenmod lmod)
1627{
1628 switch (lmod) {
1629 case OPLM_QWORD:
1630 return "";
1631 case OPLM_DWORD:
1632 return "";
1633 case OPLM_WORD:
1634 return "(u16)";
1635 case OPLM_BYTE:
1636 return "(u8)";
1637 default:
1638 ferr(po, "invalid lmod: %d\n", lmod);
1639 return "(_invalid_)";
1640 }
1641}
1642
1643static const char *lmod_cast_u_ptr(struct parsed_op *po,
1644 enum opr_lenmod lmod)
1645{
1646 switch (lmod) {
1647 case OPLM_QWORD:
1648 return "*(u64 *)";
1649 case OPLM_DWORD:
1650 return "*(u32 *)";
1651 case OPLM_WORD:
1652 return "*(u16 *)";
1653 case OPLM_BYTE:
1654 return "*(u8 *)";
1655 default:
1656 ferr(po, "invalid lmod: %d\n", lmod);
1657 return "(_invalid_)";
1658 }
1659}
1660
1661static const char *lmod_cast_s(struct parsed_op *po,
1662 enum opr_lenmod lmod)
1663{
1664 switch (lmod) {
1665 case OPLM_QWORD:
1666 return "(s64)";
1667 case OPLM_DWORD:
1668 return "(s32)";
1669 case OPLM_WORD:
1670 return "(s16)";
1671 case OPLM_BYTE:
1672 return "(s8)";
1673 default:
1674 ferr(po, "%s: invalid lmod: %d\n", __func__, lmod);
1675 return "(_invalid_)";
1676 }
1677}
1678
1679static const char *lmod_cast(struct parsed_op *po,
1680 enum opr_lenmod lmod, int is_signed)
1681{
1682 return is_signed ?
1683 lmod_cast_s(po, lmod) :
1684 lmod_cast_u(po, lmod);
1685}
1686
1687static int lmod_bytes(struct parsed_op *po, enum opr_lenmod lmod)
1688{
1689 switch (lmod) {
1690 case OPLM_QWORD:
1691 return 8;
1692 case OPLM_DWORD:
1693 return 4;
1694 case OPLM_WORD:
1695 return 2;
1696 case OPLM_BYTE:
1697 return 1;
1698 default:
1699 ferr(po, "%s: invalid lmod: %d\n", __func__, lmod);
1700 return 0;
1701 }
1702}
1703
1704static const char *opr_name(struct parsed_op *po, int opr_num)
1705{
1706 if (opr_num >= po->operand_cnt)
1707 ferr(po, "opr OOR: %d/%d\n", opr_num, po->operand_cnt);
1708 return po->operand[opr_num].name;
1709}
1710
1711static unsigned int opr_const(struct parsed_op *po, int opr_num)
1712{
1713 if (opr_num >= po->operand_cnt)
1714 ferr(po, "opr OOR: %d/%d\n", opr_num, po->operand_cnt);
1715 if (po->operand[opr_num].type != OPT_CONST)
1716 ferr(po, "opr %d: const expected\n", opr_num);
1717 return po->operand[opr_num].val;
1718}
1719
1720static const char *opr_reg_p(struct parsed_op *po, struct parsed_opr *popr)
1721{
1722 if ((unsigned int)popr->reg >= ARRAY_SIZE(regs_r32))
1723 ferr(po, "invalid reg: %d\n", popr->reg);
1724 return regs_r32[popr->reg];
1725}
1726
1727static int check_simple_cast(const char *cast, int *bits, int *is_signed)
1728{
1729 if (IS_START(cast, "(s8)") || IS_START(cast, "(u8)"))
1730 *bits = 8;
1731 else if (IS_START(cast, "(s16)") || IS_START(cast, "(u16)"))
1732 *bits = 16;
1733 else if (IS_START(cast, "(s32)") || IS_START(cast, "(u32)"))
1734 *bits = 32;
1735 else if (IS_START(cast, "(s64)") || IS_START(cast, "(u64)"))
1736 *bits = 64;
1737 else
1738 return -1;
1739
1740 *is_signed = cast[1] == 's' ? 1 : 0;
1741 return 0;
1742}
1743
1744static int check_deref_cast(const char *cast, int *bits)
1745{
1746 if (IS_START(cast, "*(u8 *)"))
1747 *bits = 8;
1748 else if (IS_START(cast, "*(u16 *)"))
1749 *bits = 16;
1750 else if (IS_START(cast, "*(u32 *)"))
1751 *bits = 32;
1752 else if (IS_START(cast, "*(u64 *)"))
1753 *bits = 64;
1754 else
1755 return -1;
1756
1757 return 0;
1758}
1759
1760// cast1 is the "final" cast
1761static const char *simplify_cast(const char *cast1, const char *cast2)
1762{
1763 static char buf[256];
1764 int bits1, bits2;
1765 int s1, s2;
1766
1767 if (cast1[0] == 0)
1768 return cast2;
1769 if (cast2[0] == 0)
1770 return cast1;
1771 if (IS(cast1, cast2))
1772 return cast1;
1773
1774 if (check_simple_cast(cast1, &bits1, &s1) == 0
1775 && check_simple_cast(cast2, &bits2, &s2) == 0)
1776 {
1777 if (bits1 <= bits2)
1778 return cast1;
1779 }
1780 if (check_simple_cast(cast1, &bits1, &s1) == 0
1781 && check_deref_cast(cast2, &bits2) == 0)
1782 {
1783 if (bits1 == bits2) {
1784 snprintf(buf, sizeof(buf), "*(%c%d *)", s1 ? 's' : 'u', bits1);
1785 return buf;
1786 }
1787 }
1788
1789 if (strchr(cast1, '*') && IS_START(cast2, "(u32)"))
1790 return cast1;
1791
1792 snprintf(buf, sizeof(buf), "%s%s", cast1, cast2);
1793 return buf;
1794}
1795
1796static const char *simplify_cast_num(const char *cast, unsigned int val)
1797{
1798 if (IS(cast, "(u8)") && val < 0x100)
1799 return "";
1800 if (IS(cast, "(s8)") && val < 0x80)
1801 return "";
1802 if (IS(cast, "(u16)") && val < 0x10000)
1803 return "";
1804 if (IS(cast, "(s16)") && val < 0x8000)
1805 return "";
1806 if (IS(cast, "(s32)") && val < 0x80000000)
1807 return "";
1808
1809 return cast;
1810}
1811
1812static struct parsed_equ *equ_find(struct parsed_op *po, const char *name,
1813 int *extra_offs)
1814{
1815 const char *p;
1816 char *endp;
1817 int namelen;
1818 int i;
1819
1820 *extra_offs = 0;
1821 namelen = strlen(name);
1822
1823 p = strpbrk(name, "+-");
1824 if (p != NULL) {
1825 namelen = p - name;
1826 if (namelen <= 0)
1827 ferr(po, "equ parse failed for '%s'\n", name);
1828
1829 errno = 0;
1830 *extra_offs = strtol(p, &endp, 16);
1831 if (*endp != 0 || errno != 0)
1832 ferr(po, "equ parse failed for '%s'\n", name);
1833 }
1834
1835 for (i = 0; i < g_eqcnt; i++)
1836 if (strncmp(g_eqs[i].name, name, namelen) == 0
1837 && g_eqs[i].name[namelen] == 0)
1838 break;
1839 if (i >= g_eqcnt) {
1840 if (po != NULL)
1841 ferr(po, "unresolved equ name: '%s'\n", name);
1842 return NULL;
1843 }
1844
1845 return &g_eqs[i];
1846}
1847
1848static int is_stack_access(struct parsed_op *po,
1849 const struct parsed_opr *popr)
1850{
1851 return (parse_stack_el(popr->name, NULL, NULL, 0)
1852 || (g_bp_frame && !(po->flags & OPF_EBP_S)
1853 && IS_START(popr->name, "ebp")));
1854}
1855
1856static void parse_stack_access(struct parsed_op *po,
1857 const char *name, char *ofs_reg, int *offset_out,
1858 int *stack_ra_out, const char **bp_arg_out, int is_lea)
1859{
1860 const char *bp_arg = "";
1861 const char *p = NULL;
1862 struct parsed_equ *eq;
1863 char *endp = NULL;
1864 int stack_ra = 0;
1865 int offset = 0;
1866
1867 ofs_reg[0] = 0;
1868
1869 if (IS_START(name, "ebp-")
1870 || (IS_START(name, "ebp+") && '0' <= name[4] && name[4] <= '9'))
1871 {
1872 p = name + 4;
1873 if (IS_START(p, "0x"))
1874 p += 2;
1875 errno = 0;
1876 offset = strtoul(p, &endp, 16);
1877 if (name[3] == '-')
1878 offset = -offset;
1879 if (*endp != 0 || errno != 0)
1880 ferr(po, "ebp- parse of '%s' failed\n", name);
1881 }
1882 else {
1883 bp_arg = parse_stack_el(name, ofs_reg, NULL, 0);
1884 eq = equ_find(po, bp_arg, &offset);
1885 if (eq == NULL)
1886 ferr(po, "detected but missing eq\n");
1887 offset += eq->offset;
1888 }
1889
1890 if (!strncmp(name, "ebp", 3))
1891 stack_ra = 4;
1892
1893 // yes it sometimes LEAs ra for compares..
1894 if (!is_lea && ofs_reg[0] == 0
1895 && stack_ra <= offset && offset < stack_ra + 4)
1896 {
1897 ferr(po, "reference to ra? %d %d\n", offset, stack_ra);
1898 }
1899
1900 *offset_out = offset;
1901 if (stack_ra_out)
1902 *stack_ra_out = stack_ra;
1903 if (bp_arg_out)
1904 *bp_arg_out = bp_arg;
1905}
1906
1907static int parse_stack_esp_offset(struct parsed_op *po,
1908 const char *name, int *offset_out)
1909{
1910 char ofs_reg[16] = { 0, };
1911 struct parsed_equ *eq;
1912 const char *bp_arg;
1913 char *endp = NULL;
1914 int base_val = 0;
1915 int offset = 0;
1916
1917 if (strstr(name, "esp") == NULL)
1918 return -1;
1919 bp_arg = parse_stack_el(name, ofs_reg, &base_val, 0);
1920 if (bp_arg == NULL) {
1921 // just plain offset?
1922 if (!IS_START(name, "esp+"))
1923 return -1;
1924 errno = 0;
1925 offset = strtol(name + 4, &endp, 0);
1926 if (endp == NULL || *endp != 0 || errno != 0)
1927 return -1;
1928 *offset_out = offset;
1929 return 0;
1930 }
1931
1932 if (ofs_reg[0] != 0)
1933 return -1;
1934 eq = equ_find(po, bp_arg, &offset);
1935 if (eq == NULL)
1936 ferr(po, "detected but missing eq\n");
1937 offset += eq->offset;
1938 *offset_out = base_val + offset;
1939 return 0;
1940}
1941
1942// returns g_func_pp arg number if arg is accessed
1943// -1 otherwise (stack vars, va_list)
1944// note: 'popr' must be from 'po', not some other op
1945static int stack_frame_access(struct parsed_op *po,
1946 struct parsed_opr *popr, char *buf, size_t buf_size,
1947 const char *name, const char *cast, int is_src, int is_lea)
1948{
1949 enum opr_lenmod tmp_lmod = OPLM_UNSPEC;
1950 const char *prefix = "";
1951 const char *bp_arg = NULL;
1952 char ofs_reg[16] = { 0, };
1953 char argname[8], buf2[32];
1954 int i, arg_i, arg_s;
1955 int unaligned = 0;
1956 int stack_ra = 0;
1957 int offset = 0;
1958 int retval = -1;
1959 int sf_ofs;
1960
1961 if (g_bp_frame && (po->flags & OPF_EBP_S)
1962 && !(po->regmask_src & mxSP))
1963 ferr(po, "stack_frame_access while ebp is scratch\n");
1964
1965 parse_stack_access(po, name, ofs_reg, &offset,
1966 &stack_ra, &bp_arg, is_lea);
1967
1968 snprintf(g_comment, sizeof(g_comment), "%s", bp_arg);
1969
1970 if (offset > stack_ra)
1971 {
1972 arg_i = (offset - stack_ra - 4) / 4;
1973 if (arg_i < 0 || arg_i >= g_func_pp->argc_stack)
1974 {
1975 if (g_func_pp->is_vararg && arg_i >= g_func_pp->argc_stack) {
1976 // vararg access - messy and non-portable,
1977 // but works with gcc on both x86 and ARM
1978 if (arg_i == g_func_pp->argc_stack)
1979 // should be va_list
1980 snprintf(buf2, sizeof(buf2), "*(u32 *)&ap");
1981 else
1982 snprintf(buf2, sizeof(buf2), "(*(u32 *)&ap + %u)",
1983 (arg_i - g_func_pp->argc_stack) * 4);
1984
1985 if (is_lea)
1986 snprintf(buf, buf_size, "%s%s", cast, buf2);
1987 else
1988 snprintf(buf, buf_size, "%s*(u32 *)%s", cast, buf2);
1989 return -1;
1990 }
1991 ferr(po, "offset 0x%x (%s,%d) doesn't map to any arg\n",
1992 offset, bp_arg, arg_i);
1993 }
1994 if (ofs_reg[0] != 0)
1995 ferr(po, "offset reg on arg access?\n");
1996
1997 for (i = arg_s = 0; i < g_func_pp->argc; i++) {
1998 if (g_func_pp->arg[i].reg != NULL)
1999 continue;
2000 if (arg_s == arg_i)
2001 break;
2002 arg_s++;
2003 }
2004 if (i == g_func_pp->argc)
2005 ferr(po, "arg %d not in prototype?\n", arg_i);
2006
2007 popr->is_ptr = g_func_pp->arg[i].type.is_ptr;
2008 retval = i;
2009
2010 snprintf(argname, sizeof(argname), "%sa%d",
2011 g_sct_func_attr & SCTFA_ARGFRAME ? "af." : "", i + 1);
2012
2013 switch (popr->lmod)
2014 {
2015 case OPLM_BYTE:
2016 if (is_lea)
2017 ferr(po, "lea/byte to arg?\n");
2018 if (is_src && (offset & 3) == 0)
2019 snprintf(buf, buf_size, "%s%s",
2020 simplify_cast(cast, "(u8)"), argname);
2021 else
2022 snprintf(buf, buf_size, "%sBYTE%d(%s)",
2023 cast, offset & 3, argname);
2024 break;
2025
2026 case OPLM_WORD:
2027 if (is_lea)
2028 ferr(po, "lea/word to arg?\n");
2029 if (offset & 1) {
2030 unaligned = 1;
2031 if (!is_src) {
2032 if (offset & 2)
2033 ferr(po, "problematic arg store\n");
2034 snprintf(buf, buf_size, "%s((char *)&%s + 1)",
2035 simplify_cast(cast, "*(u16 *)"), argname);
2036 }
2037 else
2038 ferr(po, "unaligned arg word load\n");
2039 }
2040 else if (is_src && (offset & 2) == 0)
2041 snprintf(buf, buf_size, "%s%s",
2042 simplify_cast(cast, "(u16)"), argname);
2043 else
2044 snprintf(buf, buf_size, "%s%sWORD(%s)",
2045 cast, (offset & 2) ? "HI" : "LO", argname);
2046 break;
2047
2048 case OPLM_DWORD:
2049 if (cast[0])
2050 prefix = cast;
2051 else if (is_src)
2052 prefix = "(u32)";
2053
2054 if (offset & 3) {
2055 unaligned = 1;
2056 if (is_lea)
2057 snprintf(buf, buf_size, "(u32)&%s + %d",
2058 argname, offset & 3);
2059 else if (!is_src)
2060 ferr(po, "unaligned arg store\n");
2061 else {
2062 // mov edx, [ebp+arg_4+2]; movsx ecx, dx
2063 snprintf(buf, buf_size, "%s(%s >> %d)",
2064 prefix, argname, (offset & 3) * 8);
2065 }
2066 }
2067 else {
2068 snprintf(buf, buf_size, "%s%s%s",
2069 prefix, is_lea ? "&" : "", argname);
2070 }
2071 break;
2072
2073 case OPLM_QWORD:
2074 ferr_assert(po, !(offset & 7));
2075 if (cast[0])
2076 prefix = cast;
2077 snprintf(buf, buf_size, "%s%s%s",
2078 prefix, is_lea ? "&" : "", argname);
2079 break;
2080
2081 default:
2082 ferr(po, "bp_arg bad lmod: %d\n", popr->lmod);
2083 }
2084
2085 if (unaligned)
2086 strcat(g_comment, " unaligned");
2087
2088 // common problem
2089 guess_lmod_from_c_type(&tmp_lmod, &g_func_pp->arg[i].type);
2090 if (tmp_lmod != OPLM_DWORD
2091 && (unaligned || (!is_src && lmod_bytes(po, tmp_lmod)
2092 < lmod_bytes(po, popr->lmod) + (offset & 3))))
2093 {
2094 ferr(po, "bp_arg arg%d/w offset %d and type '%s' is too small\n",
2095 i + 1, offset, g_func_pp->arg[i].type.name);
2096 }
2097 // can't check this because msvc likes to reuse
2098 // arg space for scratch..
2099 //if (popr->is_ptr && popr->lmod != OPLM_DWORD)
2100 // ferr(po, "bp_arg arg%d: non-dword ptr access\n", i + 1);
2101 }
2102 else
2103 {
2104 if (g_stack_fsz == 0)
2105 ferr(po, "stack var access without stackframe\n");
2106 g_stack_frame_used = 1;
2107
2108 sf_ofs = g_stack_fsz + offset;
2109 if (ofs_reg[0] == 0 && (offset > 0 || sf_ofs < 0))
2110 ferr(po, "bp_stack offset %d/%d\n", offset, g_stack_fsz);
2111
2112 if (is_lea)
2113 prefix = "(u32)&";
2114 else
2115 prefix = cast;
2116
2117 switch (popr->lmod)
2118 {
2119 case OPLM_BYTE:
2120 snprintf(buf, buf_size, "%ssf.b[%d%s%s]",
2121 prefix, sf_ofs, ofs_reg[0] ? "+" : "", ofs_reg);
2122 break;
2123
2124 case OPLM_WORD:
2125 if ((sf_ofs & 1) || ofs_reg[0] != 0) {
2126 // known unaligned or possibly unaligned
2127 strcat(g_comment, " unaligned");
2128 if (prefix[0] == 0)
2129 prefix = "*(u16 *)&";
2130 snprintf(buf, buf_size, "%ssf.b[%d%s%s]",
2131 prefix, sf_ofs, ofs_reg[0] ? "+" : "", ofs_reg);
2132 break;
2133 }
2134 snprintf(buf, buf_size, "%ssf.w[%d]", prefix, sf_ofs / 2);
2135 break;
2136
2137 case OPLM_DWORD:
2138 if ((sf_ofs & 3) || ofs_reg[0] != 0) {
2139 // known unaligned or possibly unaligned
2140 strcat(g_comment, " unaligned");
2141 if (prefix[0] == 0)
2142 prefix = "*(u32 *)&";
2143 snprintf(buf, buf_size, "%ssf.b[%d%s%s]",
2144 prefix, sf_ofs, ofs_reg[0] ? "+" : "", ofs_reg);
2145 break;
2146 }
2147 snprintf(buf, buf_size, "%ssf.d[%d]", prefix, sf_ofs / 4);
2148 break;
2149
2150 case OPLM_QWORD:
2151 ferr_assert(po, !(sf_ofs & 7));
2152 ferr_assert(po, ofs_reg[0] == 0);
2153 // only used for x87 int64/float, float sets is_lea
2154 if (!is_lea && (po->flags & OPF_FINT))
2155 prefix = "*(s64 *)&";
2156 snprintf(buf, buf_size, "%ssf.q[%d]", prefix, sf_ofs / 8);
2157 break;
2158
2159 default:
2160 ferr(po, "bp_stack bad lmod: %d\n", popr->lmod);
2161 }
2162 }
2163
2164 return retval;
2165}
2166
2167static void check_func_pp(struct parsed_op *po,
2168 const struct parsed_proto *pp, const char *pfx)
2169{
2170 enum opr_lenmod tmp_lmod;
2171 char buf[256];
2172 int ret, i;
2173
2174 if (pp->argc_reg != 0) {
2175 if (!g_allow_user_icall && !pp->is_fastcall) {
2176 pp_print(buf, sizeof(buf), pp);
2177 ferr(po, "%s: unexpected reg arg in icall: %s\n", pfx, buf);
2178 }
2179 if (pp->argc_stack > 0 && pp->argc_reg != 2)
2180 ferr(po, "%s: %d reg arg(s) with %d stack arg(s)\n",
2181 pfx, pp->argc_reg, pp->argc_stack);
2182 }
2183
2184 // fptrs must use 32bit args, callsite might have no information and
2185 // lack a cast to smaller types, which results in incorrectly masked
2186 // args passed (callee may assume masked args, it does on ARM)
2187 if (!pp->is_osinc) {
2188 for (i = 0; i < pp->argc; i++) {
2189 ret = guess_lmod_from_c_type(&tmp_lmod, &pp->arg[i].type);
2190 if (ret && tmp_lmod != OPLM_DWORD)
2191 ferr(po, "reference to %s with arg%d '%s'\n", pp->name,
2192 i + 1, pp->arg[i].type.name);
2193 }
2194 }
2195}
2196
2197static const char *check_label_read_ref(struct parsed_op *po,
2198 const char *name, int *is_import)
2199{
2200 const struct parsed_proto *pp;
2201
2202 pp = proto_parse(g_fhdr, name, 0);
2203 if (pp == NULL)
2204 ferr(po, "proto_parse failed for ref '%s'\n", name);
2205
2206 if (pp->is_func)
2207 check_func_pp(po, pp, "ref");
2208
2209 if (is_import != NULL)
2210 *is_import = pp->is_import;
2211
2212 return pp->name;
2213}
2214
2215static void check_opr(struct parsed_op *po, struct parsed_opr *popr)
2216{
2217 if (popr->segment == SEG_FS)
2218 ferr(po, "fs: used\n");
2219 if (popr->segment == SEG_GS)
2220 ferr(po, "gs: used\n");
2221}
2222
2223static char *out_src_opr(char *buf, size_t buf_size,
2224 struct parsed_op *po, struct parsed_opr *popr, const char *cast,
2225 int is_lea)
2226{
2227 char tmp1[256], tmp2[256];
2228 char expr[256];
2229 const char *name;
2230 int is_import = 0;
2231 char *p;
2232 int ret;
2233
2234 check_opr(po, popr);
2235
2236 if (cast == NULL)
2237 cast = "";
2238
2239 switch (popr->type) {
2240 case OPT_REG:
2241 if (is_lea)
2242 ferr(po, "lea from reg?\n");
2243
2244 switch (popr->lmod) {
2245 case OPLM_QWORD:
2246 snprintf(buf, buf_size, "%s%s.q", cast, opr_reg_p(po, popr));
2247 break;
2248 case OPLM_DWORD:
2249 snprintf(buf, buf_size, "%s%s", cast, opr_reg_p(po, popr));
2250 break;
2251 case OPLM_WORD:
2252 snprintf(buf, buf_size, "%s%s",
2253 simplify_cast(cast, "(u16)"), opr_reg_p(po, popr));
2254 break;
2255 case OPLM_BYTE:
2256 if (popr->name[1] == 'h') // XXX..
2257 snprintf(buf, buf_size, "%s(%s >> 8)",
2258 simplify_cast(cast, "(u8)"), opr_reg_p(po, popr));
2259 else
2260 snprintf(buf, buf_size, "%s%s",
2261 simplify_cast(cast, "(u8)"), opr_reg_p(po, popr));
2262 break;
2263 default:
2264 ferr(po, "invalid src lmod: %d\n", popr->lmod);
2265 }
2266 break;
2267
2268 case OPT_REGMEM:
2269 if (is_stack_access(po, popr)) {
2270 stack_frame_access(po, popr, buf, buf_size,
2271 popr->name, cast, 1, is_lea);
2272 break;
2273 }
2274
2275 strcpy(expr, popr->name);
2276 if (strchr(expr, '[')) {
2277 // special case: '[' can only be left for label[reg] form
2278 ret = sscanf(expr, "%[^[][%[^]]]", tmp1, tmp2);
2279 if (ret != 2)
2280 ferr(po, "parse failure for '%s'\n", expr);
2281 if (tmp1[0] == '(') {
2282 // (off_4FFF50+3)[eax]
2283 p = strchr(tmp1 + 1, ')');
2284 if (p == NULL || p[1] != 0)
2285 ferr(po, "parse failure (2) for '%s'\n", expr);
2286 *p = 0;
2287 memmove(tmp1, tmp1 + 1, strlen(tmp1));
2288 }
2289 snprintf(expr, sizeof(expr), "(u32)&%s + %s", tmp1, tmp2);
2290 }
2291
2292 // XXX: do we need more parsing?
2293 if (is_lea) {
2294 snprintf(buf, buf_size, "%s", expr);
2295 break;
2296 }
2297
2298 snprintf(buf, buf_size, "%s(%s)",
2299 simplify_cast(cast, lmod_cast_u_ptr(po, popr->lmod)), expr);
2300 break;
2301
2302 case OPT_LABEL:
2303 name = check_label_read_ref(po, popr->name, &is_import);
2304 if (is_import)
2305 // for imported data, asm is loading the offset
2306 goto do_offset;
2307
2308 if (cast[0] == 0 && popr->is_ptr)
2309 cast = "(u32)";
2310
2311 if (is_lea)
2312 snprintf(buf, buf_size, "(u32)&%s", name);
2313 else if (popr->size_lt)
2314 snprintf(buf, buf_size, "%s%s%s%s", cast,
2315 lmod_cast_u_ptr(po, popr->lmod),
2316 popr->is_array ? "" : "&", name);
2317 else
2318 snprintf(buf, buf_size, "%s%s%s", cast, name,
2319 popr->is_array ? "[0]" : "");
2320 break;
2321
2322 case OPT_OFFSET:
2323 do_offset:
2324 name = check_label_read_ref(po, popr->name, NULL);
2325 if (cast[0] == 0)
2326 cast = "(u32)";
2327 if (is_lea)
2328 ferr(po, "lea an offset?\n");
2329 snprintf(buf, buf_size, "%s&%s", cast, name);
2330 break;
2331
2332 case OPT_CONST:
2333 if (is_lea)
2334 ferr(po, "lea from const?\n");
2335
2336 printf_number(tmp1, sizeof(tmp1), popr->val);
2337 if (popr->val == 0 && strchr(cast, '*'))
2338 snprintf(buf, buf_size, "NULL");
2339 else
2340 snprintf(buf, buf_size, "%s%s",
2341 simplify_cast_num(cast, popr->val), tmp1);
2342 break;
2343
2344 default:
2345 ferr(po, "invalid src type: %d\n", popr->type);
2346 }
2347
2348 return buf;
2349}
2350
2351// note: may set is_ptr (we find that out late for ebp frame..)
2352static char *out_dst_opr(char *buf, size_t buf_size,
2353 struct parsed_op *po, struct parsed_opr *popr)
2354{
2355 check_opr(po, popr);
2356
2357 switch (popr->type) {
2358 case OPT_REG:
2359 switch (popr->lmod) {
2360 case OPLM_QWORD:
2361 snprintf(buf, buf_size, "%s.q", opr_reg_p(po, popr));
2362 break;
2363 case OPLM_DWORD:
2364 snprintf(buf, buf_size, "%s", opr_reg_p(po, popr));
2365 break;
2366 case OPLM_WORD:
2367 // ugh..
2368 snprintf(buf, buf_size, "LOWORD(%s)", opr_reg_p(po, popr));
2369 break;
2370 case OPLM_BYTE:
2371 // ugh..
2372 if (popr->name[1] == 'h') // XXX..
2373 snprintf(buf, buf_size, "BYTE1(%s)", opr_reg_p(po, popr));
2374 else
2375 snprintf(buf, buf_size, "LOBYTE(%s)", opr_reg_p(po, popr));
2376 break;
2377 default:
2378 ferr(po, "invalid dst lmod: %d\n", popr->lmod);
2379 }
2380 break;
2381
2382 case OPT_REGMEM:
2383 if (is_stack_access(po, popr)) {
2384 stack_frame_access(po, popr, buf, buf_size,
2385 popr->name, "", 0, 0);
2386 break;
2387 }
2388
2389 return out_src_opr(buf, buf_size, po, popr, NULL, 0);
2390
2391 case OPT_LABEL:
2392 if (popr->size_mismatch)
2393 snprintf(buf, buf_size, "%s%s%s",
2394 lmod_cast_u_ptr(po, popr->lmod),
2395 popr->is_array ? "" : "&", popr->name);
2396 else
2397 snprintf(buf, buf_size, "%s%s", popr->name,
2398 popr->is_array ? "[0]" : "");
2399 break;
2400
2401 default:
2402 ferr(po, "invalid dst type: %d\n", popr->type);
2403 }
2404
2405 return buf;
2406}
2407
2408static char *out_src_opr_u32(char *buf, size_t buf_size,
2409 struct parsed_op *po, struct parsed_opr *popr)
2410{
2411 return out_src_opr(buf, buf_size, po, popr, NULL, 0);
2412}
2413
2414// do we need a helper func to perform a float i/o?
2415static int float_opr_needs_helper(struct parsed_op *po,
2416 struct parsed_opr *popr)
2417{
2418 if (!(g_sct_func_attr & SCTFA_UA_FLOAT))
2419 return 0;
2420 if (popr->type != OPT_REGMEM)
2421 return 0;
2422 if (is_stack_access(po, popr))
2423 return 0;
2424
2425 return 1;
2426}
2427
2428static char *out_opr_float(char *buf, size_t buf_size,
2429 struct parsed_op *po, struct parsed_opr *popr, int is_src,
2430 int need_float_stack)
2431{
2432 const char *cast = NULL;
2433 char tmp[256];
2434 union {
2435 float f;
2436 int i;
2437 } u;
2438
2439 switch (popr->type) {
2440 case OPT_REG:
2441 if (popr->reg < xST0 || popr->reg > xST7) {
2442 // func arg
2443 ferr_assert(po, po->op == OP_PUSH);
2444 ferr_assert(po, popr->lmod == OPLM_DWORD);
2445 snprintf(buf, buf_size, "*(float *)&%s", opr_reg_p(po, popr));
2446 break;
2447 }
2448
2449 if (need_float_stack) {
2450 if (popr->reg == xST0)
2451 snprintf(buf, buf_size, "f_st[f_stp & 7]");
2452 else
2453 snprintf(buf, buf_size, "f_st[(f_stp + %d) & 7]",
2454 popr->reg - xST0);
2455 }
2456 else
2457 snprintf(buf, buf_size, "f_st%d", popr->reg - xST0);
2458 break;
2459
2460 case OPT_REGMEM:
2461 if (popr->lmod == OPLM_QWORD && is_stack_access(po, popr)) {
2462 stack_frame_access(po, popr, buf, buf_size,
2463 popr->name, "", is_src, 0);
2464 break;
2465 }
2466 // fallthrough
2467 case OPT_LABEL:
2468 case OPT_OFFSET:
2469 switch (popr->lmod) {
2470 case OPLM_QWORD:
2471 cast = "double";
2472 break;
2473 case OPLM_DWORD:
2474 cast = "float";
2475 break;
2476 default:
2477 ferr(po, "unhandled lmod: %d\n", popr->lmod);
2478 break;
2479 }
2480 out_src_opr(tmp, sizeof(tmp), po, popr, "", 1);
2481 if (is_src && float_opr_needs_helper(po, popr))
2482 snprintf(buf, buf_size, "%s_load(%s)", cast, tmp);
2483 else
2484 snprintf(buf, buf_size, "*(%s *)(%s)", cast, tmp);
2485 break;
2486
2487 case OPT_CONST:
2488 // only for func float args pushes
2489 ferr_assert(po, po->op == OP_PUSH);
2490 u.i = po->operand[0].val;
2491 if (ceilf(u.f) == u.f)
2492 snprintf(buf, buf_size, "%.1ff", u.f);
2493 else
2494 snprintf(buf, buf_size, "%.8ff", u.f);
2495 break;
2496
2497 default:
2498 ferr(po, "invalid float type: %d\n", popr->type);
2499 }
2500
2501 return buf;
2502}
2503
2504static char *out_src_opr_float(char *buf, size_t buf_size,
2505 struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
2506{
2507 return out_opr_float(buf, buf_size, po, popr, 1, need_float_stack);
2508}
2509
2510static char *out_dst_opr_float(char *buf, size_t buf_size,
2511 struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
2512{
2513 return out_opr_float(buf, buf_size, po, popr, 0, need_float_stack);
2514}
2515
2516static void out_test_for_cc(char *buf, size_t buf_size,
2517 struct parsed_op *po, enum parsed_flag_op pfo, int is_inv,
2518 enum opr_lenmod lmod, const char *expr)
2519{
2520 const char *cast, *scast;
2521
2522 cast = lmod_cast_u(po, lmod);
2523 scast = lmod_cast_s(po, lmod);
2524
2525 switch (pfo) {
2526 case PFO_Z:
2527 case PFO_BE: // CF==1||ZF==1; CF=0
2528 snprintf(buf, buf_size, "(%s%s %s 0)",
2529 cast, expr, is_inv ? "!=" : "==");
2530 break;
2531
2532 case PFO_S:
2533 case PFO_L: // SF!=OF; OF=0
2534 snprintf(buf, buf_size, "(%s%s %s 0)",
2535 scast, expr, is_inv ? ">=" : "<");
2536 break;
2537
2538 case PFO_LE: // ZF==1||SF!=OF; OF=0
2539 snprintf(buf, buf_size, "(%s%s %s 0)",
2540 scast, expr, is_inv ? ">" : "<=");
2541 break;
2542
2543 case PFO_C: // CF=0
2544 case PFO_O: // OF=0
2545 snprintf(buf, buf_size, "(%d)", !!is_inv);
2546 break;
2547
2548 case PFO_P: // PF==1
2549 snprintf(buf, buf_size, "(%sdo_parity(%s))",
2550 is_inv ? "!" : "", expr);
2551 break;
2552
2553 default:
2554 ferr(po, "%s: unhandled parsed_flag_op: %d\n", __func__, pfo);
2555 }
2556}
2557
2558static void out_cmp_for_cc(char *buf, size_t buf_size,
2559 struct parsed_op *po, enum parsed_flag_op pfo, int is_inv,
2560 int is_neg)
2561{
2562 const char *cast, *scast, *cast_use;
2563 char buf1[256], buf2[256];
2564 enum opr_lenmod lmod;
2565
2566 if (po->op != OP_DEC && po->operand[0].lmod != po->operand[1].lmod)
2567 ferr(po, "%s: lmod mismatch: %d %d\n", __func__,
2568 po->operand[0].lmod, po->operand[1].lmod);
2569 lmod = po->operand[0].lmod;
2570
2571 cast = lmod_cast_u(po, lmod);
2572 scast = lmod_cast_s(po, lmod);
2573
2574 switch (pfo) {
2575 case PFO_C:
2576 case PFO_Z:
2577 case PFO_BE: // !a
2578 cast_use = cast;
2579 break;
2580
2581 case PFO_S:
2582 case PFO_L: // !ge
2583 case PFO_LE:
2584 cast_use = scast;
2585 break;
2586
2587 default:
2588 ferr(po, "%s: unhandled parsed_flag_op: %d\n", __func__, pfo);
2589 }
2590
2591 out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], cast_use, 0);
2592 if (po->op == OP_DEC)
2593 snprintf(buf2, sizeof(buf2), "1");
2594 else {
2595 char cast_op2[64];
2596 snprintf(cast_op2, sizeof(cast_op2) - 1, "%s", cast_use);
2597 if (is_neg)
2598 strcat(cast_op2, "-");
2599 out_src_opr(buf2, sizeof(buf2), po, &po->operand[1], cast_op2, 0);
2600 }
2601
2602 switch (pfo) {
2603 case PFO_C:
2604 // note: must be unsigned compare
2605 snprintf(buf, buf_size, "(%s %s %s)",
2606 buf1, is_inv ? ">=" : "<", buf2);
2607 break;
2608
2609 case PFO_Z:
2610 snprintf(buf, buf_size, "(%s %s %s)",
2611 buf1, is_inv ? "!=" : "==", buf2);
2612 break;
2613
2614 case PFO_BE: // !a
2615 // note: must be unsigned compare
2616 snprintf(buf, buf_size, "(%s %s %s)",
2617 buf1, is_inv ? ">" : "<=", buf2);
2618
2619 // annoying case
2620 if (is_inv && lmod == OPLM_BYTE
2621 && po->operand[1].type == OPT_CONST
2622 && po->operand[1].val == 0xff)
2623 {
2624 snprintf(g_comment, sizeof(g_comment), "if %s", buf);
2625 snprintf(buf, buf_size, "(0)");
2626 }
2627 break;
2628
2629 // note: must be signed compare
2630 case PFO_S:
2631 snprintf(buf, buf_size, "(%s(%s - %s) %s 0)",
2632 scast, buf1, buf2, is_inv ? ">=" : "<");
2633 break;
2634
2635 case PFO_L: // !ge
2636 snprintf(buf, buf_size, "(%s %s %s)",
2637 buf1, is_inv ? ">=" : "<", buf2);
2638 break;
2639
2640 case PFO_LE: // !g
2641 snprintf(buf, buf_size, "(%s %s %s)",
2642 buf1, is_inv ? ">" : "<=", buf2);
2643 break;
2644
2645 default:
2646 break;
2647 }
2648}
2649
2650static void out_cmp_test(char *buf, size_t buf_size,
2651 struct parsed_op *po, enum parsed_flag_op pfo, int is_inv)
2652{
2653 char buf1[256], buf2[256], buf3[256];
2654
2655 if (po->op == OP_TEST) {
2656 if (IS(opr_name(po, 0), opr_name(po, 1))) {
2657 out_src_opr_u32(buf3, sizeof(buf3), po, &po->operand[0]);
2658 }
2659 else {
2660 out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
2661 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
2662 snprintf(buf3, sizeof(buf3), "(%s & %s)", buf1, buf2);
2663 }
2664 out_test_for_cc(buf, buf_size, po, pfo, is_inv,
2665 po->operand[0].lmod, buf3);
2666 }
2667 else if (po->op == OP_CMP) {
2668 out_cmp_for_cc(buf, buf_size, po, pfo, is_inv, 0);
2669 }
2670 else
2671 ferr(po, "%s: unhandled op: %d\n", __func__, po->op);
2672}
2673
2674static void propagate_lmod(struct parsed_op *po, struct parsed_opr *popr1,
2675 struct parsed_opr *popr2)
2676{
2677 if (popr1->lmod == OPLM_UNSPEC && popr2->lmod == OPLM_UNSPEC)
2678 ferr(po, "missing lmod for both operands\n");
2679
2680 if (popr1->lmod == OPLM_UNSPEC)
2681 popr1->lmod = popr2->lmod;
2682 else if (popr2->lmod == OPLM_UNSPEC)
2683 popr2->lmod = popr1->lmod;
2684 else if (popr1->lmod != popr2->lmod) {
2685 if (popr1->type_from_var) {
2686 popr1->size_mismatch = 1;
2687 if (popr1->lmod < popr2->lmod)
2688 popr1->size_lt = 1;
2689 popr1->lmod = popr2->lmod;
2690 }
2691 else if (popr2->type_from_var) {
2692 popr2->size_mismatch = 1;
2693 if (popr2->lmod < popr1->lmod)
2694 popr2->size_lt = 1;
2695 popr2->lmod = popr1->lmod;
2696 }
2697 else
2698 ferr(po, "conflicting lmods: %d vs %d\n",
2699 popr1->lmod, popr2->lmod);
2700 }
2701}
2702
2703static const char *op_to_c(struct parsed_op *po)
2704{
2705 switch (po->op)
2706 {
2707 case OP_ADD:
2708 case OP_ADC:
2709 return "+";
2710 case OP_SUB:
2711 case OP_SBB:
2712 return "-";
2713 case OP_AND:
2714 return "&";
2715 case OP_OR:
2716 return "|";
2717 case OP_XOR:
2718 return "^";
2719 case OP_SHL:
2720 return "<<";
2721 case OP_SHR:
2722 return ">>";
2723 case OP_MUL:
2724 case OP_IMUL:
2725 return "*";
2726 default:
2727 ferr(po, "op_to_c was supplied with %d\n", po->op);
2728 }
2729}
2730
2731// last op in stream - unconditional branch or ret
2732#define LAST_OP(_i) ((ops[_i].flags & OPF_TAIL) \
2733 || ((ops[_i].flags & (OPF_JMP|OPF_CJMP|OPF_RMD)) == OPF_JMP \
2734 && ops[_i].op != OP_CALL))
2735
2736#define check_i(po, i) \
2737 if ((i) < 0) \
2738 ferr(po, "bad " #i ": %d\n", i)
2739
2740// note: this skips over calls and rm'd stuff assuming they're handled
2741// so it's intended to use at one of final passes
2742// exception: doesn't skip OPF_RSAVE stuff
2743static int scan_for_pop(int i, int opcnt, int magic, int reg,
2744 int depth, int seen_noreturn, int save_level, int flags_set)
2745{
2746 struct parsed_op *po;
2747 int relevant;
2748 int ret = 0;
2749 int j;
2750
2751 for (; i < opcnt; i++) {
2752 po = &ops[i];
2753 if (po->cc_scratch == magic)
2754 return ret; // already checked
2755 po->cc_scratch = magic;
2756
2757 if (po->flags & OPF_TAIL) {
2758 if (po->op == OP_CALL && po->pp != NULL && po->pp->is_noreturn) {
2759 // msvc sometimes generates stack cleanup code after
2760 // noreturn, set a flag and continue
2761 seen_noreturn = 1;
2762
2763 // ... but stop if there is another path to next insn -
2764 // if msvc skipped something stack tracking may mess up
2765 if (i + 1 < opcnt && g_labels[i + 1] != NULL)
2766 goto out;
2767 }
2768 else
2769 goto out;
2770 }
2771
2772 if (po->flags & OPF_FARG)
2773 continue;
2774 if (po->flags & (OPF_RMD|OPF_DONE)) {
2775 if (!(po->flags & OPF_RSAVE))
2776 continue;
2777 // reprocess, there might be another push in some "parallel"
2778 // path that took a pop what we should also take
2779 }
2780
2781 if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
2782 if (po->btj != NULL) {
2783 // jumptable
2784 for (j = 0; j < po->btj->count; j++) {
2785 check_i(po, po->btj->d[j].bt_i);
2786 ret |= scan_for_pop(po->btj->d[j].bt_i, opcnt, magic, reg,
2787 depth, seen_noreturn, save_level, flags_set);
2788 if (ret < 0)
2789 return ret; // dead end
2790 }
2791 return ret;
2792 }
2793
2794 check_i(po, po->bt_i);
2795 if (po->flags & OPF_CJMP) {
2796 ret |= scan_for_pop(po->bt_i, opcnt, magic, reg,
2797 depth, seen_noreturn, save_level, flags_set);
2798 if (ret < 0)
2799 return ret; // dead end
2800 }
2801 else {
2802 i = po->bt_i - 1;
2803 }
2804 continue;
2805 }
2806
2807 relevant = 0;
2808 if ((po->op == OP_POP || po->op == OP_PUSH)
2809 && po->operand[0].type == OPT_REG && po->operand[0].reg == reg)
2810 {
2811 relevant = 1;
2812 }
2813
2814 if (po->op == OP_PUSH) {
2815 depth++;
2816 }
2817 else if (po->op == OP_POP) {
2818 if (relevant && depth == 0) {
2819 if (flags_set == 0 && save_level > 0) {
2820 ret = scan_for_pop(i + 1, opcnt, magic, reg,
2821 depth, seen_noreturn, save_level - 1, flags_set);
2822 if (ret != 1)
2823 // no pop for other levels, current one must be false
2824 return -1;
2825 }
2826 po->flags |= flags_set;
2827 return 1;
2828 }
2829 depth--;
2830 }
2831 }
2832
2833out:
2834 // for noreturn, assume msvc skipped stack cleanup
2835 return seen_noreturn ? 1 : -1;
2836}
2837
2838// scan for 'reg' pop backwards starting from i
2839// intended to use for register restore search, so other reg
2840// references are considered an error
2841static int scan_for_rsave_pop_reg(int i, int magic, int reg, int set_flags)
2842{
2843 struct parsed_op *po;
2844 struct label_ref *lr;
2845 int ret = 0;
2846
2847 ops[i].cc_scratch = magic;
2848
2849 while (1)
2850 {
2851 if (g_labels[i] != NULL) {
2852 lr = &g_label_refs[i];
2853 for (; lr != NULL; lr = lr->next) {
2854 check_i(&ops[i], lr->i);
2855 ret |= scan_for_rsave_pop_reg(lr->i, magic, reg, set_flags);
2856 if (ret < 0)
2857 return ret;
2858 }
2859 if (i > 0 && LAST_OP(i - 1))
2860 return ret;
2861 }
2862
2863 i--;
2864 if (i < 0)
2865 break;
2866
2867 if (ops[i].cc_scratch == magic)
2868 return ret;
2869 ops[i].cc_scratch = magic;
2870
2871 po = &ops[i];
2872 if (po->op == OP_POP && po->operand[0].reg == reg) {
2873 if (po->flags & (OPF_RMD|OPF_DONE))
2874 return -1;
2875
2876 po->flags |= set_flags;
2877 return 1;
2878 }
2879
2880 // this also covers the case where we reach corresponding push
2881 if ((po->regmask_dst | po->regmask_src) & (1 << reg))
2882 return -1;
2883 }
2884
2885 // nothing interesting on this path,
2886 // still return ret for something recursive calls could find
2887 return ret;
2888}
2889
2890static void find_reachable_exits(int i, int opcnt, int magic,
2891 int *exits, int *exit_count)
2892{
2893 struct parsed_op *po;
2894 int j;
2895
2896 for (; i < opcnt; i++)
2897 {
2898 po = &ops[i];
2899 if (po->cc_scratch == magic)
2900 return;
2901 po->cc_scratch = magic;
2902
2903 if (po->flags & OPF_TAIL) {
2904 ferr_assert(po, *exit_count < MAX_EXITS);
2905 exits[*exit_count] = i;
2906 (*exit_count)++;
2907 return;
2908 }
2909
2910 if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
2911 if (po->flags & OPF_RMD)
2912 continue;
2913
2914 if (po->btj != NULL) {
2915 for (j = 0; j < po->btj->count; j++) {
2916 check_i(po, po->btj->d[j].bt_i);
2917 find_reachable_exits(po->btj->d[j].bt_i, opcnt, magic,
2918 exits, exit_count);
2919 }
2920 return;
2921 }
2922
2923 check_i(po, po->bt_i);
2924 if (po->flags & OPF_CJMP)
2925 find_reachable_exits(po->bt_i, opcnt, magic, exits, exit_count);
2926 else
2927 i = po->bt_i - 1;
2928 continue;
2929 }
2930 }
2931}
2932
2933// scan for 'reg' pop backwards starting from exits (all paths)
2934static int scan_for_pop_ret(int i, int opcnt, int reg, int set_flags)
2935{
2936 static int exits[MAX_EXITS];
2937 static int exit_count;
2938 int found = 0;
2939 int e, j, ret;
2940
2941 if (!set_flags) {
2942 exit_count = 0;
2943 find_reachable_exits(i, opcnt, i + opcnt * 15, exits,
2944 &exit_count);
2945 ferr_assert(&ops[i], exit_count > 0);
2946 }
2947
2948 for (j = 0; j < exit_count; j++) {
2949 e = exits[j];
2950 ret = scan_for_rsave_pop_reg(e, i + opcnt * 16 + set_flags,
2951 reg, set_flags);
2952 if (ret != -1) {
2953 found |= ret;
2954 continue;
2955 }
2956 if (ops[e].op == OP_CALL && ops[e].pp != NULL
2957 && ops[e].pp->is_noreturn)
2958 {
2959 // assume stack cleanup was skipped
2960 continue;
2961 }
2962 return -1;
2963 }
2964
2965 return found;
2966}
2967
2968// scan for one or more pop of push <const>
2969static int scan_for_pop_const_r(int i, int opcnt, int magic,
2970 int push_i, int is_probe)
2971{
2972 struct parsed_op *po;
2973 struct label_ref *lr;
2974 int ret = 0;
2975 int j;
2976
2977 for (; i < opcnt; i++)
2978 {
2979 po = &ops[i];
2980 if (po->cc_scratch == magic)
2981 return ret; // already checked
2982 po->cc_scratch = magic;
2983
2984 if (po->flags & OPF_JMP) {
2985 if (po->flags & OPF_RMD)
2986 continue;
2987 if (po->op == OP_CALL)
2988 return -1;
2989
2990 if (po->btj != NULL) {
2991 for (j = 0; j < po->btj->count; j++) {
2992 check_i(po, po->btj->d[j].bt_i);
2993 ret |= scan_for_pop_const_r(po->btj->d[j].bt_i, opcnt, magic,
2994 push_i, is_probe);
2995 if (ret < 0)
2996 return ret;
2997 }
2998 return ret;
2999 }
3000
3001 check_i(po, po->bt_i);
3002 if (po->flags & OPF_CJMP) {
3003 ret |= scan_for_pop_const_r(po->bt_i, opcnt, magic, push_i,
3004 is_probe);
3005 if (ret < 0)
3006 return ret;
3007 }
3008 else {
3009 i = po->bt_i - 1;
3010 }
3011 continue;
3012 }
3013
3014 if ((po->flags & (OPF_TAIL|OPF_RSAVE)) || po->op == OP_PUSH)
3015 return -1;
3016
3017 if (g_labels[i] != NULL) {
3018 // all refs must be visited
3019 lr = &g_label_refs[i];
3020 for (; lr != NULL; lr = lr->next) {
3021 check_i(po, lr->i);
3022 if (ops[lr->i].cc_scratch != magic)
3023 return -1;
3024 }
3025 if (i > 0 && !LAST_OP(i - 1) && ops[i - 1].cc_scratch != magic)
3026 return -1;
3027 }
3028
3029 if (po->op == OP_POP)
3030 {
3031 if (po->flags & (OPF_RMD|OPF_DONE))
3032 return -1;
3033
3034 if (!is_probe) {
3035 po->flags |= OPF_DONE;
3036 po->datap = &ops[push_i];
3037 }
3038 return 1;
3039 }
3040 }
3041
3042 return -1;
3043}
3044
3045static void scan_for_pop_const(int i, int opcnt, int magic)
3046{
3047 int ret;
3048
3049 ret = scan_for_pop_const_r(i + 1, opcnt, magic, i, 1);
3050 if (ret == 1) {
3051 ops[i].flags |= OPF_RMD | OPF_DONE;
3052 scan_for_pop_const_r(i + 1, opcnt, magic + 1, i, 0);
3053 }
3054}
3055
3056// check if all branch targets within a marked path are also marked
3057// note: the path checked must not be empty or end with a branch
3058static int check_path_branches(int opcnt, int magic)
3059{
3060 struct parsed_op *po;
3061 int i, j;
3062
3063 for (i = 0; i < opcnt; i++) {
3064 po = &ops[i];
3065 if (po->cc_scratch != magic)
3066 continue;
3067
3068 if (po->flags & OPF_JMP) {
3069 if ((po->flags & OPF_RMD) || po->op == OP_CALL)
3070 continue;
3071
3072 if (po->btj != NULL) {
3073 for (j = 0; j < po->btj->count; j++) {
3074 check_i(po, po->btj->d[j].bt_i);
3075 if (ops[po->btj->d[j].bt_i].cc_scratch != magic)
3076 return 0;
3077 }
3078 }
3079
3080 check_i(po, po->bt_i);
3081 if (ops[po->bt_i].cc_scratch != magic)
3082 return 0;
3083 if ((po->flags & OPF_CJMP) && ops[i + 1].cc_scratch != magic)
3084 return 0;
3085 }
3086 }
3087
3088 return 1;
3089}
3090
3091// scan for multiple pushes for given pop
3092static int scan_pushes_for_pop_r(int i, int magic, int pop_i,
3093 int is_probe)
3094{
3095 int reg = ops[pop_i].operand[0].reg;
3096 struct parsed_op *po;
3097 struct label_ref *lr;
3098 int ret = 0;
3099
3100 ops[i].cc_scratch = magic;
3101
3102 while (1)
3103 {
3104 if (g_labels[i] != NULL) {
3105 lr = &g_label_refs[i];
3106 for (; lr != NULL; lr = lr->next) {
3107 check_i(&ops[i], lr->i);
3108 ret |= scan_pushes_for_pop_r(lr->i, magic, pop_i, is_probe);
3109 if (ret < 0)
3110 return ret;
3111 }
3112 if (i > 0 && LAST_OP(i - 1))
3113 return ret;
3114 }
3115
3116 i--;
3117 if (i < 0)
3118 break;
3119
3120 if (ops[i].cc_scratch == magic)
3121 return ret;
3122 ops[i].cc_scratch = magic;
3123
3124 po = &ops[i];
3125 if (po->op == OP_CALL)
3126 return -1;
3127 if ((po->flags & (OPF_TAIL|OPF_RSAVE)) || po->op == OP_POP)
3128 return -1;
3129
3130 if (po->op == OP_PUSH)
3131 {
3132 if (po->datap != NULL)
3133 return -1;
3134 if (po->operand[0].type == OPT_REG && po->operand[0].reg == reg)
3135 // leave this case for reg save/restore handlers
3136 return -1;
3137
3138 if (!is_probe) {
3139 po->flags |= OPF_PPUSH | OPF_DONE;
3140 po->datap = &ops[pop_i];
3141 }
3142 return 1;
3143 }
3144 }
3145
3146 return -1;
3147}
3148
3149static void scan_pushes_for_pop(int i, int opcnt, int *regmask_pp)
3150{
3151 int magic = i + opcnt * 14;
3152 int ret;
3153
3154 ret = scan_pushes_for_pop_r(i, magic, i, 1);
3155 if (ret == 1) {
3156 ret = check_path_branches(opcnt, magic);
3157 if (ret == 1) {
3158 ops[i].flags |= OPF_PPUSH | OPF_DONE;
3159 *regmask_pp |= 1 << ops[i].operand[0].reg;
3160 scan_pushes_for_pop_r(i, magic + 1, i, 0);
3161 }
3162 }
3163}
3164
3165static void scan_propagate_df(int i, int opcnt)
3166{
3167 struct parsed_op *po = &ops[i];
3168 int j;
3169
3170 for (; i < opcnt; i++) {
3171 po = &ops[i];
3172 if (po->flags & OPF_DF)
3173 return; // already resolved
3174 po->flags |= OPF_DF;
3175
3176 if (po->op == OP_CALL)
3177 ferr(po, "call with DF set?\n");
3178
3179 if (po->flags & OPF_JMP) {
3180 if (po->btj != NULL) {
3181 // jumptable
3182 for (j = 0; j < po->btj->count; j++) {
3183 check_i(po, po->btj->d[j].bt_i);
3184 scan_propagate_df(po->btj->d[j].bt_i, opcnt);
3185 }
3186 return;
3187 }
3188
3189 if (po->flags & OPF_RMD)
3190 continue;
3191 check_i(po, po->bt_i);
3192 if (po->flags & OPF_CJMP)
3193 scan_propagate_df(po->bt_i, opcnt);
3194 else
3195 i = po->bt_i - 1;
3196 continue;
3197 }
3198
3199 if (po->flags & OPF_TAIL)
3200 break;
3201
3202 if (po->op == OP_CLD) {
3203 po->flags |= OPF_RMD | OPF_DONE;
3204 return;
3205 }
3206 }
3207
3208 ferr(po, "missing DF clear?\n");
3209}
3210
3211// is operand 'opr' referenced by parsed_op 'po'?
3212static int is_opr_referenced(const struct parsed_opr *opr,
3213 const struct parsed_op *po)
3214{
3215 int i, mask;
3216
3217 if (opr->type == OPT_REG) {
3218 mask = po->regmask_dst | po->regmask_src;
3219 if (po->op == OP_CALL)
3220 mask |= (1 << xAX) | (1 << xCX) | (1 << xDX);
3221 if ((1 << opr->reg) & mask)
3222 return 1;
3223 else
3224 return 0;
3225 }
3226
3227 for (i = 0; i < po->operand_cnt; i++)
3228 if (IS(po->operand[0].name, opr->name))
3229 return 1;
3230
3231 return 0;
3232}
3233
3234// is operand 'opr' read by parsed_op 'po'?
3235static int is_opr_read(const struct parsed_opr *opr,
3236 const struct parsed_op *po)
3237{
3238 if (opr->type == OPT_REG) {
3239 if (po->regmask_src & (1 << opr->reg))
3240 return 1;
3241 else
3242 return 0;
3243 }
3244
3245 // yes I'm lazy
3246 return 0;
3247}
3248
3249// is operand 'opr' modified by parsed_op 'po'?
3250static int is_opr_modified(const struct parsed_opr *opr,
3251 const struct parsed_op *po)
3252{
3253 int mask;
3254
3255 if (opr->type == OPT_REG) {
3256 if (po->op == OP_CALL) {
3257 mask = po->regmask_dst;
3258 mask |= (1 << xAX) | (1 << xCX) | (1 << xDX); // ?
3259 if (mask & (1 << opr->reg))
3260 return 1;
3261 else
3262 return 0;
3263 }
3264
3265 if (po->regmask_dst & (1 << opr->reg))
3266 return 1;
3267 else
3268 return 0;
3269 }
3270
3271 return IS(po->operand[0].name, opr->name);
3272}
3273
3274// is any operand of parsed_op 'po_test' modified by parsed_op 'po'?
3275static int is_any_opr_modified(const struct parsed_op *po_test,
3276 const struct parsed_op *po, int c_mode)
3277{
3278 int mask;
3279 int i;
3280
3281 if ((po->flags & OPF_RMD) || !(po->flags & OPF_DATA))
3282 return 0;
3283
3284 if (po_test->operand_cnt == 1 && po_test->operand[0].type == OPT_CONST)
3285 return 0;
3286
3287 if ((po_test->regmask_src | po_test->regmask_dst) & po->regmask_dst)
3288 return 1;
3289
3290 // in reality, it can wreck any register, but in decompiled C
3291 // version it can only overwrite eax or edx:eax
3292 mask = (1 << xAX) | (1 << xDX);
3293 if (!c_mode)
3294 mask |= 1 << xCX;
3295
3296 if (po->op == OP_CALL
3297 && ((po_test->regmask_src | po_test->regmask_dst) & mask))
3298 return 1;
3299
3300 for (i = 0; i < po_test->operand_cnt; i++)
3301 if (IS(po_test->operand[i].name, po->operand[0].name))
3302 return 1;
3303
3304 return 0;
3305}
3306
3307// scan for any po_test operand modification in range given
3308static int scan_for_mod(struct parsed_op *po_test, int i, int opcnt,
3309 int c_mode)
3310{
3311 if (po_test->operand_cnt == 1 && po_test->operand[0].type == OPT_CONST)
3312 return -1;
3313
3314 for (; i < opcnt; i++) {
3315 if (is_any_opr_modified(po_test, &ops[i], c_mode))
3316 return i;
3317 }
3318
3319 return -1;
3320}
3321
3322// scan for po_test operand[0] modification in range given
3323static int scan_for_mod_opr0(struct parsed_op *po_test,
3324 int i, int opcnt)
3325{
3326 for (; i < opcnt; i++) {
3327 if (is_opr_modified(&po_test->operand[0], &ops[i]))
3328 return i;
3329 }
3330
3331 return -1;
3332}
3333
3334static int try_resolve_const(int i, const struct parsed_opr *opr,
3335 int magic, unsigned int *val);
3336
3337static int scan_for_flag_set(int i, int opcnt, int magic,
3338 int *branched, int *setters, int *setter_cnt)
3339{
3340 struct label_ref *lr;
3341 int ret;
3342
3343 while (i >= 0) {
3344 if (ops[i].cc_scratch == magic) {
3345 // is this a problem?
3346 //ferr(&ops[i], "%s looped\n", __func__);
3347 return 0;
3348 }
3349 ops[i].cc_scratch = magic;
3350
3351 if (g_labels[i] != NULL) {
3352 *branched = 1;
3353
3354 lr = &g_label_refs[i];
3355 for (; lr->next; lr = lr->next) {
3356 check_i(&ops[i], lr->i);
3357 ret = scan_for_flag_set(lr->i, opcnt, magic,
3358 branched, setters, setter_cnt);
3359 if (ret < 0)
3360 return ret;
3361 }
3362
3363 check_i(&ops[i], lr->i);
3364 if (i > 0 && LAST_OP(i - 1)) {
3365 i = lr->i;
3366 continue;
3367 }
3368 ret = scan_for_flag_set(lr->i, opcnt, magic,
3369 branched, setters, setter_cnt);
3370 if (ret < 0)
3371 return ret;
3372 }
3373 i--;
3374
3375 if (ops[i].flags & OPF_FLAGS) {
3376 setters[*setter_cnt] = i;
3377 (*setter_cnt)++;
3378
3379 if (ops[i].flags & OPF_REP) {
3380 struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, xCX);
3381 unsigned int uval;
3382
3383 ret = try_resolve_const(i, &opr, i + opcnt * 7, &uval);
3384 if (ret != 1 || uval == 0) {
3385 // can't treat it as full setter because of ecx=0 case,
3386 // also disallow delayed compare
3387 *branched = 1;
3388 continue;
3389 }
3390 }
3391
3392 return 0;
3393 }
3394
3395 if ((ops[i].flags & (OPF_JMP|OPF_CJMP)) == OPF_JMP)
3396 return -1;
3397 }
3398
3399 return -1;
3400}
3401
3402// scan back for cdq, if anything modifies edx, fail
3403static int scan_for_cdq_edx(int i)
3404{
3405 while (i >= 0) {
3406 if (g_labels[i] != NULL) {
3407 if (g_label_refs[i].next != NULL)
3408 return -1;
3409 if (i > 0 && LAST_OP(i - 1)) {
3410 i = g_label_refs[i].i;
3411 continue;
3412 }
3413 return -1;
3414 }
3415 i--;
3416
3417 if (ops[i].op == OP_CDQ)
3418 return i;
3419
3420 if (ops[i].regmask_dst & (1 << xDX))
3421 return -1;
3422 }
3423
3424 return -1;
3425}
3426
3427static int scan_for_reg_clear(int i, int reg)
3428{
3429 while (i >= 0) {
3430 if (g_labels[i] != NULL) {
3431 if (g_label_refs[i].next != NULL)
3432 return -1;
3433 if (i > 0 && LAST_OP(i - 1)) {
3434 i = g_label_refs[i].i;
3435 continue;
3436 }
3437 return -1;
3438 }
3439 i--;
3440
3441 if (ops[i].op == OP_XOR
3442 && ops[i].operand[0].lmod == OPLM_DWORD
3443 && ops[i].operand[0].reg == ops[i].operand[1].reg
3444 && ops[i].operand[0].reg == reg)
3445 return i;
3446
3447 if (ops[i].regmask_dst & (1 << reg))
3448 return -1;
3449 }
3450
3451 return -1;
3452}
3453
3454static void patch_esp_adjust(struct parsed_op *po, int adj)
3455{
3456 ferr_assert(po, po->op == OP_ADD);
3457 ferr_assert(po, IS(opr_name(po, 0), "esp"));
3458 ferr_assert(po, po->operand[1].type == OPT_CONST);
3459
3460 // this is a bit of a hack, but deals with use of
3461 // single adj for multiple calls
3462 po->operand[1].val -= adj;
3463 po->flags |= OPF_RMD;
3464 if (po->operand[1].val == 0)
3465 po->flags |= OPF_DONE;
3466 ferr_assert(po, (int)po->operand[1].val >= 0);
3467}
3468
3469// scan for positive, constant esp adjust
3470// multipath case is preliminary
3471static int scan_for_esp_adjust(int i, int opcnt,
3472 int adj_expect, int *adj, int *is_multipath, int do_update)
3473{
3474 int adj_expect_unknown = 0;
3475 struct parsed_op *po;
3476 int first_pop = -1;
3477 int adj_best = 0;
3478
3479 *adj = *is_multipath = 0;
3480 if (adj_expect < 0) {
3481 adj_expect_unknown = 1;
3482 adj_expect = 32 * 4; // enough?
3483 }
3484
3485 for (; i < opcnt && *adj < adj_expect; i++) {
3486 if (g_labels[i] != NULL)
3487 *is_multipath = 1;
3488
3489 po = &ops[i];
3490 if (po->flags & OPF_DONE)
3491 continue;
3492
3493 if (po->op == OP_ADD && po->operand[0].reg == xSP) {
3494 if (po->operand[1].type != OPT_CONST)
3495 ferr(&ops[i], "non-const esp adjust?\n");
3496 *adj += po->operand[1].val;
3497 if (*adj & 3)
3498 ferr(&ops[i], "unaligned esp adjust: %x\n", *adj);
3499 if (do_update) {
3500 if (!*is_multipath)
3501 patch_esp_adjust(po, adj_expect);
3502 else
3503 po->flags |= OPF_RMD;
3504 }
3505 return i;
3506 }
3507 else if (po->op == OP_PUSH) {
3508 //if (first_pop == -1)
3509 // first_pop = -2; // none
3510 *adj -= lmod_bytes(po, po->operand[0].lmod);
3511 }
3512 else if (po->op == OP_POP) {
3513 if (!(po->flags & OPF_DONE)) {
3514 // seems like msvc only uses 'pop ecx' for stack realignment..
3515 if (po->operand[0].type != OPT_REG || po->operand[0].reg != xCX)
3516 break;
3517 if (first_pop == -1 && *adj >= 0)
3518 first_pop = i;
3519 }
3520 if (do_update && *adj >= 0) {
3521 po->flags |= OPF_RMD;
3522 if (!*is_multipath)
3523 po->flags |= OPF_DONE | OPF_NOREGS;
3524 }
3525
3526 *adj += lmod_bytes(po, po->operand[0].lmod);
3527 if (*adj > adj_best)
3528 adj_best = *adj;
3529 }
3530 else if (po->flags & (OPF_JMP|OPF_TAIL)) {
3531 if (po->op == OP_JMP && po->btj == NULL) {
3532 if (po->bt_i <= i)
3533 break;
3534 i = po->bt_i - 1;
3535 continue;
3536 }
3537 if (po->op != OP_CALL)
3538 break;
3539 if (po->operand[0].type != OPT_LABEL)
3540 break;
3541 if (po->pp != NULL && po->pp->is_stdcall)
3542 break;
3543 if (adj_expect_unknown && first_pop >= 0)
3544 break;
3545 // assume it's another cdecl call
3546 }
3547 }
3548
3549 if (first_pop >= 0) {
3550 // probably only 'pop ecx' was used
3551 *adj = adj_best;
3552 return first_pop;
3553 }
3554
3555 return -1;
3556}
3557
3558static void scan_fwd_set_flags(int i, int opcnt, int magic, int flags)
3559{
3560 struct parsed_op *po;
3561 int j;
3562
3563 if (i < 0)
3564 ferr(ops, "%s: followed bad branch?\n", __func__);
3565
3566 for (; i < opcnt; i++) {
3567 po = &ops[i];
3568 if (po->cc_scratch == magic)
3569 return;
3570 po->cc_scratch = magic;
3571 po->flags |= flags;
3572
3573 if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
3574 if (po->btj != NULL) {
3575 // jumptable
3576 for (j = 0; j < po->btj->count; j++)
3577 scan_fwd_set_flags(po->btj->d[j].bt_i, opcnt, magic, flags);
3578 return;
3579 }
3580
3581 scan_fwd_set_flags(po->bt_i, opcnt, magic, flags);
3582 if (!(po->flags & OPF_CJMP))
3583 return;
3584 }
3585 if (po->flags & OPF_TAIL)
3586 return;
3587 }
3588}
3589
3590static const struct parsed_proto *try_recover_pp(
3591 struct parsed_op *po, const struct parsed_opr *opr,
3592 int is_call, int *search_instead)
3593{
3594 const struct parsed_proto *pp = NULL;
3595 char buf[256];
3596 char *p;
3597
3598 if (po->pp != NULL && (po->flags & OPF_DATA)) {
3599 // hint given in asm
3600 return po->pp;
3601 }
3602
3603 // maybe an arg of g_func?
3604 if (opr->type == OPT_REGMEM && is_stack_access(po, opr))
3605 {
3606 char ofs_reg[16] = { 0, };
3607 int arg, arg_s, arg_i;
3608 int stack_ra = 0;
3609 int offset = 0;
3610
3611 if (g_header_mode)
3612 return NULL;
3613
3614 parse_stack_access(po, opr->name, ofs_reg,
3615 &offset, &stack_ra, NULL, 0);
3616 if (ofs_reg[0] != 0)
3617 ferr(po, "offset reg on arg access?\n");
3618 if (offset <= stack_ra) {
3619 // search who set the stack var instead
3620 if (search_instead != NULL)
3621 *search_instead = 1;
3622 return NULL;
3623 }
3624
3625 arg_i = (offset - stack_ra - 4) / 4;
3626 for (arg = arg_s = 0; arg < g_func_pp->argc; arg++) {
3627 if (g_func_pp->arg[arg].reg != NULL)
3628 continue;
3629 if (arg_s == arg_i)
3630 break;
3631 arg_s++;
3632 }
3633 if (arg == g_func_pp->argc)
3634 ferr(po, "stack arg %d not in prototype?\n", arg_i);
3635
3636 pp = g_func_pp->arg[arg].pp;
3637 if (is_call) {
3638 if (pp == NULL)
3639 ferr(po, "icall arg: arg%d has no pp\n", arg + 1);
3640 check_func_pp(po, pp, "icall arg");
3641 }
3642 }
3643 else if (opr->type == OPT_REGMEM && strchr(opr->name + 1, '[')) {
3644 // label[index]
3645 p = strchr(opr->name + 1, '[');
3646 memcpy(buf, opr->name, p - opr->name);
3647 buf[p - opr->name] = 0;
3648 pp = proto_parse(g_fhdr, buf, g_quiet_pp);
3649 }
3650 else if (opr->type == OPT_OFFSET || opr->type == OPT_LABEL) {
3651 pp = proto_parse(g_fhdr, opr->name, g_quiet_pp);
3652 if (pp == NULL) {
3653 if (!g_header_mode)
3654 ferr(po, "proto_parse failed for icall to '%s'\n", opr->name);
3655 }
3656 else
3657 check_func_pp(po, pp, "reg-fptr ref");
3658 }
3659
3660 return pp;
3661}
3662
3663static void scan_for_call_type(int i, const struct parsed_opr *opr,
3664 int magic, int is_call_op, const struct parsed_proto **pp_found,
3665 int *pp_i, int *multi)
3666{
3667 const struct parsed_proto *pp = NULL;
3668 struct parsed_op *po;
3669 struct label_ref *lr;
3670
3671 ops[i].cc_scratch = magic;
3672
3673 while (1) {
3674 if (g_labels[i] != NULL) {
3675 lr = &g_label_refs[i];
3676 for (; lr != NULL; lr = lr->next) {
3677 check_i(&ops[i], lr->i);
3678 scan_for_call_type(lr->i, opr, magic, is_call_op,
3679 pp_found, pp_i, multi);
3680 }
3681 if (i > 0 && LAST_OP(i - 1))
3682 return;
3683 }
3684
3685 i--;
3686 if (i < 0)
3687 break;
3688
3689 if (ops[i].cc_scratch == magic)
3690 return;
3691 ops[i].cc_scratch = magic;
3692
3693 if (!(ops[i].flags & OPF_DATA))
3694 continue;
3695 if (!is_opr_modified(opr, &ops[i]))
3696 continue;
3697 if (ops[i].op != OP_MOV && ops[i].op != OP_LEA) {
3698 // most probably trashed by some processing
3699 *pp_found = NULL;
3700 return;
3701 }
3702
3703 opr = &ops[i].operand[1];
3704 if (opr->type != OPT_REG)
3705 break;
3706 }
3707
3708 po = (i >= 0) ? &ops[i] : ops;
3709
3710 if (i < 0) {
3711 // reached the top - can only be an arg-reg
3712 if (opr->type != OPT_REG || g_func_pp == NULL)
3713 return;
3714
3715 for (i = 0; i < g_func_pp->argc; i++) {
3716 if (g_func_pp->arg[i].reg == NULL)
3717 continue;
3718 if (IS(opr->name, g_func_pp->arg[i].reg))
3719 break;
3720 }
3721 if (i == g_func_pp->argc)
3722 return;
3723 pp = g_func_pp->arg[i].pp;
3724 if (pp == NULL) {
3725 if (is_call_op)
3726 ferr(po, "icall: arg%d (%s) is not a fptr?\n",
3727 i + 1, g_func_pp->arg[i].reg);
3728 return;
3729 }
3730 check_func_pp(po, pp, "icall reg-arg");
3731 }
3732 else
3733 pp = try_recover_pp(po, opr, is_call_op, NULL);
3734
3735 if (*pp_found != NULL && pp != NULL && *pp_found != pp) {
3736 if (pp_cmp_func(*pp_found, pp)) {
3737 if (pp_i != NULL && *pp_i != -1)
3738 fnote(&ops[*pp_i], "(other ref)\n");
3739 ferr(po, "icall: parsed_proto mismatch\n");
3740 }
3741 if (multi != NULL)
3742 *multi = 1;
3743 }
3744 if (pp != NULL) {
3745 *pp_found = pp;
3746 if (pp_i != NULL)
3747 *pp_i = po - ops;
3748 }
3749}
3750
3751static void add_label_ref(struct label_ref *lr, int op_i)
3752{
3753 struct label_ref *lr_new;
3754
3755 if (lr->i == -1) {
3756 lr->i = op_i;
3757 return;
3758 }
3759
3760 lr_new = calloc(1, sizeof(*lr_new));
3761 lr_new->i = op_i;
3762 lr_new->next = lr->next;
3763 lr->next = lr_new;
3764}
3765
3766static struct parsed_data *try_resolve_jumptab(int i, int opcnt)
3767{
3768 struct parsed_op *po = &ops[i];
3769 struct parsed_data *pd;
3770 char label[NAMELEN], *p;
3771 int len, j, l;
3772
3773 p = strchr(po->operand[0].name, '[');
3774 if (p == NULL)
3775 return NULL;
3776
3777 len = p - po->operand[0].name;
3778 strncpy(label, po->operand[0].name, len);
3779 label[len] = 0;
3780
3781 for (j = 0, pd = NULL; j < g_func_pd_cnt; j++) {
3782 if (IS(g_func_pd[j].label, label)) {
3783 pd = &g_func_pd[j];
3784 break;
3785 }
3786 }
3787 if (pd == NULL)
3788 //ferr(po, "label '%s' not parsed?\n", label);
3789 return NULL;
3790
3791 if (pd->type != OPT_OFFSET)
3792 ferr(po, "label '%s' with non-offset data?\n", label);
3793
3794 // find all labels, link
3795 for (j = 0; j < pd->count; j++) {
3796 for (l = 0; l < opcnt; l++) {
3797 if (g_labels[l] != NULL && IS(g_labels[l], pd->d[j].u.label)) {
3798 add_label_ref(&g_label_refs[l], i);
3799 pd->d[j].bt_i = l;
3800 break;
3801 }
3802 }
3803 }
3804
3805 return pd;
3806}
3807
3808static void clear_labels(int count)
3809{
3810 int i;
3811
3812 for (i = 0; i < count; i++) {
3813 if (g_labels[i] != NULL) {
3814 free(g_labels[i]);
3815 g_labels[i] = NULL;
3816 }
3817 }
3818}
3819
3820static int get_pp_arg_regmask_src(const struct parsed_proto *pp)
3821{
3822 int regmask = 0;
3823 int i, reg;
3824
3825 for (i = 0; i < pp->argc; i++) {
3826 if (pp->arg[i].reg != NULL) {
3827 reg = char_array_i(regs_r32,
3828 ARRAY_SIZE(regs_r32), pp->arg[i].reg);
3829 if (reg < 0)
3830 ferr(ops, "arg '%s' of func '%s' is not a reg?\n",
3831 pp->arg[i].reg, pp->name);
3832 regmask |= 1 << reg;
3833 }
3834 }
3835
3836 return regmask;
3837}
3838
3839static int get_pp_arg_regmask_dst(const struct parsed_proto *pp)
3840{
3841 int regmask = 0;
3842 int i, reg;
3843
3844 if (pp->has_retreg) {
3845 for (i = 0; i < pp->argc; i++) {
3846 if (pp->arg[i].type.is_retreg) {
3847 reg = char_array_i(regs_r32,
3848 ARRAY_SIZE(regs_r32), pp->arg[i].reg);
3849 ferr_assert(ops, reg >= 0);
3850 regmask |= 1 << reg;
3851 }
3852 }
3853 }
3854
3855 if (strstr(pp->ret_type.name, "int64"))
3856 return regmask | (1 << xAX) | (1 << xDX);
3857 if (IS(pp->ret_type.name, "float")
3858 || IS(pp->ret_type.name, "double"))
3859 {
3860 return regmask | mxST0;
3861 }
3862 if (strcasecmp(pp->ret_type.name, "void") == 0)
3863 return regmask;
3864
3865 return regmask | mxAX;
3866}
3867
3868static int are_ops_same(struct parsed_op *po1, struct parsed_op *po2)
3869{
3870 return po1->op == po2->op && po1->operand_cnt == po2->operand_cnt
3871 && memcmp(po1->operand, po2->operand,
3872 sizeof(po1->operand[0]) * po1->operand_cnt) == 0;
3873}
3874
3875static void resolve_branches_parse_calls(int opcnt)
3876{
3877 static const struct {
3878 const char *name;
3879 enum op_op op;
3880 unsigned int flags;
3881 unsigned int regmask_src;
3882 unsigned int regmask_dst;
3883 } pseudo_ops[] = {
3884 { "__allshl", OPP_ALLSHL, OPF_DATA, mxAX|mxDX|mxCX, mxAX|mxDX },
3885 { "__allshr", OPP_ALLSHR, OPF_DATA, mxAX|mxDX|mxCX, mxAX|mxDX },
3886 { "__ftol", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
3887 // more precise? Wine gets away with just __ftol handler
3888 { "__ftol2", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
3889 { "__CIpow", OPP_CIPOW, OPF_FPOP, mxST0|mxST1, mxST0 },
3890 };
3891 const struct parsed_proto *pp_c;
3892 struct parsed_proto *pp;
3893 struct parsed_data *pd;
3894 struct parsed_op *po;
3895 const char *tmpname;
3896 enum op_op prev_op;
3897 int i, l;
3898 int ret;
3899
3900 for (i = 0; i < opcnt; i++)
3901 {
3902 po = &ops[i];
3903 po->bt_i = -1;
3904 po->btj = NULL;
3905
3906 if (po->datap != NULL) {
3907 pp = calloc(1, sizeof(*pp));
3908 my_assert_not(pp, NULL);
3909
3910 ret = parse_protostr(po->datap, pp);
3911 if (ret < 0)
3912 ferr(po, "bad protostr supplied: %s\n", (char *)po->datap);
3913 free(po->datap);
3914 po->datap = NULL;
3915 po->pp = pp;
3916 }
3917
3918 if (po->op == OP_CALL) {
3919 pp = NULL;
3920
3921 if (po->pp != NULL)
3922 pp = po->pp;
3923 else if (po->operand[0].type == OPT_LABEL)
3924 {
3925 tmpname = opr_name(po, 0);
3926 if (IS_START(tmpname, "loc_")) {
3927 if (!g_seh_found)
3928 ferr(po, "call to loc_*\n");
3929 // eliminate_seh() must take care of it
3930 continue;
3931 }
3932 if (IS(tmpname, "__alloca_probe"))
3933 continue;
3934 if (IS(tmpname, "__SEH_prolog")) {
3935 ferr_assert(po, g_seh_found == 0);
3936 g_seh_found = 2;
3937 continue;
3938 }
3939 if (IS(tmpname, "__SEH_epilog"))
3940 continue;
3941
3942 // convert some calls to pseudo-ops
3943 for (l = 0; l < ARRAY_SIZE(pseudo_ops); l++) {
3944 if (!IS(tmpname, pseudo_ops[l].name))
3945 continue;
3946
3947 po->op = pseudo_ops[l].op;
3948 po->operand_cnt = 0;
3949 po->regmask_src = pseudo_ops[l].regmask_src;
3950 po->regmask_dst = pseudo_ops[l].regmask_dst;
3951 po->flags &= OPF_TAIL;
3952 po->flags |= pseudo_ops[l].flags;
3953 po->flags |= po->regmask_dst ? OPF_DATA : 0;
3954 break;
3955 }
3956 if (l < ARRAY_SIZE(pseudo_ops))
3957 continue;
3958
3959 pp_c = proto_parse(g_fhdr, tmpname, g_header_mode);
3960 if (!g_header_mode && pp_c == NULL)
3961 ferr(po, "proto_parse failed for call '%s'\n", tmpname);
3962
3963 if (pp_c != NULL) {
3964 pp = proto_clone(pp_c);
3965 my_assert_not(pp, NULL);
3966 }
3967 }
3968
3969 if (pp != NULL) {
3970 if (pp->is_fptr)
3971 check_func_pp(po, pp, "fptr var call");
3972 if (pp->is_noreturn) {
3973 po->flags |= OPF_TAIL;
3974 po->flags &= ~OPF_ATAIL; // most likely...
3975 }
3976 }
3977 po->pp = pp;
3978 continue;
3979 }
3980
3981 if (!(po->flags & OPF_JMP) || po->op == OP_RET)
3982 continue;
3983
3984 if (po->operand[0].type == OPT_REGMEM) {
3985 pd = try_resolve_jumptab(i, opcnt);
3986 if (pd == NULL)
3987 goto tailcall;
3988
3989 po->btj = pd;
3990 continue;
3991 }
3992
3993 for (l = 0; l < opcnt; l++) {
3994 if (g_labels[l] != NULL
3995 && IS(po->operand[0].name, g_labels[l]))
3996 {
3997 if (l == i + 1 && po->op == OP_JMP) {
3998 // yet another alignment type...
3999 po->flags |= OPF_RMD | OPF_DONE;
4000 po->flags &= ~OPF_JMP;
4001 po->op = OP_NOP;
4002 break;
4003 }
4004 add_label_ref(&g_label_refs[l], i);
4005 po->bt_i = l;
4006 break;
4007 }
4008 }
4009
4010 if (po->bt_i != -1 || (po->flags & OPF_RMD))
4011 continue;
4012
4013 if (po->operand[0].type == OPT_LABEL
4014 || po->operand[0].type == OPT_REG)
4015 // assume tail call
4016 goto tailcall;
4017
4018 ferr(po, "unhandled branch\n");
4019
4020tailcall:
4021 po->op = OP_CALL;
4022 po->flags |= OPF_TAIL;
4023 prev_op = i > 0 ? ops[i - 1].op : OP_UD2;
4024 if (prev_op == OP_POP)
4025 po->flags |= OPF_ATAIL;
4026 if (g_stack_fsz + g_bp_frame == 0 && prev_op != OP_PUSH
4027 && (g_func_pp == NULL || g_func_pp->argc_stack > 0))
4028 {
4029 po->flags |= OPF_ATAIL;
4030 }
4031 i--; // reprocess
4032 }
4033}
4034
4035static int resolve_origin(int i, const struct parsed_opr *opr,
4036 int magic, int *op_i, int *is_caller);
4037static void set_label(int i, const char *name);
4038
4039static void eliminate_seh_writes(int opcnt)
4040{
4041 const struct parsed_opr *opr;
4042 char ofs_reg[16];
4043 int offset;
4044 int i;
4045
4046 // assume all sf writes above g_seh_size to be seh related
4047 // (probably unsafe but oh well)
4048 for (i = 0; i < opcnt; i++) {
4049 if (ops[i].op != OP_MOV)
4050 continue;
4051 opr = &ops[i].operand[0];
4052 if (opr->type != OPT_REGMEM)
4053 continue;
4054 if (!is_stack_access(&ops[i], opr))
4055 continue;
4056
4057 offset = 0;
4058 parse_stack_access(&ops[i], opr->name, ofs_reg, &offset,
4059 NULL, NULL, 0);
4060 if (offset < 0 && offset >= -g_seh_size)
4061 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4062 }
4063}
4064
4065static void eliminate_seh_finally(int opcnt)
4066{
4067 const char *target_name = NULL;
4068 const char *return_name = NULL;
4069 int exits[MAX_EXITS];
4070 int exit_count = 0;
4071 int call_i = -1;
4072 int target_i = -1;
4073 int return_i = -1;
4074 int tgend_i = -1;
4075 int i;
4076
4077 for (i = 0; i < opcnt; i++) {
4078 if (ops[i].op != OP_CALL)
4079 continue;
4080 if (!IS_START(opr_name(&ops[i], 0), "loc_"))
4081 continue;
4082 if (target_name != NULL)
4083 ferr(&ops[i], "multiple finally calls? (last was %s)\n",
4084 target_name);
4085 target_name = opr_name(&ops[i], 0);
4086 call_i = i;
4087
4088 if (g_labels[i + 1] == NULL)
4089 set_label(i + 1, "seh_fin_done");
4090 return_name = g_labels[i + 1];
4091 return_i = i + 1;
4092 }
4093
4094 if (call_i == -1)
4095 // no finally block
4096 return;
4097
4098 // find finally code (bt_i is not set because it's call)
4099 for (i = 0; i < opcnt; i++) {
4100 if (g_labels[i] == NULL)
4101 continue;
4102 if (!IS(g_labels[i], target_name))
4103 continue;
4104
4105 ferr_assert(&ops[i], target_i == -1);
4106 target_i = i;
4107 }
4108 ferr_assert(&ops[0], target_i != -1);
4109
4110 find_reachable_exits(target_i, opcnt, target_i + opcnt * 24,
4111 exits, &exit_count);
4112 ferr_assert(&ops[target_i], exit_count == 1);
4113 ferr_assert(&ops[target_i], ops[exits[0]].op == OP_RET);
4114 tgend_i = exits[0];
4115
4116 // convert to jumps, link
4117 ops[call_i].op = OP_JMP;
4118 ops[call_i].bt_i = target_i;
4119 add_label_ref(&g_label_refs[target_i], call_i);
4120
4121 ops[tgend_i].op = OP_JMP;
4122 ops[tgend_i].flags &= ~OPF_TAIL;
4123 ops[tgend_i].flags |= OPF_JMP;
4124 ops[tgend_i].bt_i = return_i;
4125 ops[tgend_i].operand_cnt = 1;
4126 ops[tgend_i].operand[0].type = OPT_LABEL;
4127 snprintf(ops[tgend_i].operand[0].name, NAMELEN, "%s", return_name);
4128 add_label_ref(&g_label_refs[return_i], tgend_i);
4129
4130 // rm seh finally entry code
4131 for (i = target_i - 1; i >= 0; i--) {
4132 if (g_labels[i] != NULL && g_label_refs[i].i != -1)
4133 return;
4134 if (ops[i].flags & OPF_CJMP)
4135 return;
4136 if (ops[i].flags & (OPF_JMP | OPF_TAIL))
4137 break;
4138 }
4139 for (i = target_i - 1; i >= 0; i--) {
4140 if (ops[i].flags & (OPF_JMP | OPF_TAIL))
4141 break;
4142 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4143 }
4144}
4145
4146static void eliminate_seh(int opcnt)
4147{
4148 int i, j, k, ret;
4149
4150 for (i = 0; i < opcnt; i++) {
4151 if (ops[i].op != OP_MOV)
4152 continue;
4153 if (ops[i].operand[0].segment != SEG_FS)
4154 continue;
4155 if (!IS(opr_name(&ops[i], 0), "0"))
4156 continue;
4157
4158 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4159 if (ops[i].operand[1].reg == xSP) {
4160 for (j = i - 1; j >= 0; j--) {
4161 if (ops[j].op != OP_PUSH)
4162 continue;
4163 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4164 g_seh_size += 4;
4165 if (ops[j].operand[0].val == ~0)
4166 break;
4167 if (ops[j].operand[0].type == OPT_REG) {
4168 k = -1;
4169 ret = resolve_origin(j, &ops[j].operand[0],
4170 j + opcnt * 22, &k, NULL);
4171 if (ret == 1)
4172 ops[k].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4173 }
4174 }
4175 if (j < 0)
4176 ferr(ops, "missing seh terminator\n");
4177 }
4178 else {
4179 k = -1;
4180 ret = resolve_origin(i, &ops[i].operand[1],
4181 i + opcnt * 23, &k, NULL);
4182 if (ret == 1)
4183 ops[k].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4184 }
4185 }
4186
4187 eliminate_seh_writes(opcnt);
4188 eliminate_seh_finally(opcnt);
4189}
4190
4191static void eliminate_seh_calls(int opcnt)
4192{
4193 int epilog_found = 0;
4194 int i;
4195
4196 g_bp_frame = 1;
4197 g_seh_size = 0x10;
4198
4199 i = 0;
4200 ferr_assert(&ops[i], ops[i].op == OP_PUSH
4201 && ops[i].operand[0].type == OPT_CONST);
4202 g_stack_fsz = g_seh_size + ops[i].operand[0].val;
4203 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4204
4205 i++;
4206 ferr_assert(&ops[i], ops[i].op == OP_PUSH
4207 && ops[i].operand[0].type == OPT_OFFSET);
4208 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4209
4210 i++;
4211 ferr_assert(&ops[i], ops[i].op == OP_CALL
4212 && IS(opr_name(&ops[i], 0), "__SEH_prolog"));
4213 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4214
4215 for (i++; i < opcnt; i++) {
4216 if (ops[i].op != OP_CALL)
4217 continue;
4218 if (!IS(opr_name(&ops[i], 0), "__SEH_epilog"))
4219 continue;
4220
4221 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4222 epilog_found = 1;
4223 }
4224 ferr_assert(ops, epilog_found);
4225
4226 eliminate_seh_writes(opcnt);
4227 eliminate_seh_finally(opcnt);
4228}
4229
4230// check for prologue of many pushes and epilogue with pops
4231static void check_simple_sequence(int opcnt, int *fsz)
4232{
4233 int found = 0;
4234 int seq_len;
4235 int seq_p;
4236 int seq[4];
4237 int reg;
4238 int i, j;
4239
4240 for (i = 0; i < opcnt && i < ARRAY_SIZE(seq); i++) {
4241 if (ops[i].op != OP_PUSH || ops[i].operand[0].type != OPT_REG)
4242 break;
4243 reg = ops[i].operand[0].reg;
4244 if (reg != xBX && reg != xSI && reg != xDI && reg != xBP)
4245 break;
4246 for (j = 0; j < i; j++)
4247 if (seq[j] == reg)
4248 break;
4249 if (j != i)
4250 // probably something else is going on here
4251 break;
4252 seq[i] = reg;
4253 }
4254 seq_len = i;
4255 if (seq_len == 0)
4256 return;
4257
4258 for (; i < opcnt && seq_len > 0; i++) {
4259 if (!(ops[i].flags & OPF_TAIL))
4260 continue;
4261
4262 for (j = i - 1, seq_p = 0; j >= 0 && seq_p < seq_len; j--) {
4263 if (ops[j].op != OP_POP || ops[j].operand[0].type != OPT_REG)
4264 break;
4265 if (ops[j].operand[0].reg != seq[seq_p])
4266 break;
4267 seq_p++;
4268 }
4269 found = seq_len = seq_p;
4270 }
4271 if (!found)
4272 return;
4273
4274 for (i = 0; i < seq_len; i++)
4275 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4276
4277 for (; i < opcnt && seq_len > 0; i++) {
4278 if (!(ops[i].flags & OPF_TAIL))
4279 continue;
4280
4281 for (j = i - 1, seq_p = 0; j >= 0 && seq_p < seq_len; j--) {
4282 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4283 seq_p++;
4284 }
4285 }
4286
4287 // unlike pushes after sub esp,
4288 // IDA treats pushes like this as part of var area
4289 *fsz += seq_len * 4;
4290}
4291
4292static int scan_prologue_ecx(int i, int opcnt, int flags_set,
4293 int limit, int *ecx_push_out)
4294{
4295 const struct parsed_proto *pp;
4296 int ecx_push = 0, other_push = 0;
4297 int ret;
4298
4299 while (limit > 0 && ops[i].op == OP_PUSH
4300 && IS(opr_name(&ops[i], 0), "ecx"))
4301 {
4302 ops[i].flags |= flags_set;
4303 ecx_push++;
4304 i++;
4305 limit--;
4306 }
4307
4308 ret = i;
4309 if (ecx_push == 0 || flags_set != 0)
4310 goto out;
4311
4312 // check if some of the pushes aren't really call args
4313 for (; i < opcnt; i++) {
4314 if (i > 0 && g_labels[i] != NULL)
4315 break;
4316 if (ops[i].flags & (OPF_JMP|OPF_TAIL))
4317 break;
4318 if (ops[i].op == OP_PUSH)
4319 other_push++;
4320 }
4321
4322 if (ops[i].op != OP_CALL)
4323 goto out;
4324
4325 pp = ops[i].pp;
4326 if (pp == NULL && ops[i].operand[0].type == OPT_LABEL)
4327 pp = proto_parse(g_fhdr, opr_name(&ops[i], 0), 1);
4328 if (pp == NULL)
4329 goto out;
4330
4331 ferr_assert(&ops[i], ecx_push + other_push >= pp->argc_stack);
4332 if (other_push < pp->argc_stack)
4333 ecx_push -= pp->argc_stack - other_push;
4334
4335out:
4336 if (ecx_push_out != NULL)
4337 *ecx_push_out = ecx_push;
4338 return ret;
4339}
4340
4341static int scan_prologue(int i, int opcnt, int *ecx_push, int *esp_sub)
4342{
4343 const char *name;
4344 int j, len, ret;
4345 int ecx_tmp = 0;
4346
4347 for (; i < opcnt; i++)
4348 if (!(ops[i].flags & OPF_DONE))
4349 break;
4350
4351 ret = scan_prologue_ecx(i, opcnt, 0, 4, &ecx_tmp);
4352 if (ecx_tmp > 0) {
4353 scan_prologue_ecx(i, opcnt, OPF_RMD | OPF_DONE | OPF_NOREGS,
4354 ecx_tmp, NULL);
4355 g_stack_fsz += 4 * ecx_tmp;
4356 *ecx_push += ecx_tmp;
4357 i = ret;
4358 }
4359
4360 for (; i < opcnt; i++) {
4361 if (i > 0 && g_labels[i] != NULL)
4362 break;
4363 if (ops[i].flags & (OPF_JMP|OPF_TAIL))
4364 break;
4365 if (ops[i].flags & OPF_DONE)
4366 continue;
4367 if (ops[i].op == OP_PUSH)
4368 break;
4369 if (ops[i].op == OP_SUB && ops[i].operand[0].reg == xSP
4370 && ops[i].operand[1].type == OPT_CONST)
4371 {
4372 g_stack_fsz += opr_const(&ops[i], 1);
4373 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4374 i++;
4375 *esp_sub = 1;
4376 break;
4377 }
4378 if (ops[i].op == OP_LEA && ops[i].operand[0].reg == xSP
4379 && ops[i].operand[1].type == OPT_REGMEM
4380 && IS_START(ops[i].operand[1].name, "esp-"))
4381 {
4382 name = ops[i].operand[1].name;
4383 ret = sscanf(name, "esp-%x%n", &j, &len);
4384 ferr_assert(&ops[i], ret == 1 && len == strlen(name));
4385 g_stack_fsz += j;
4386 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4387 i++;
4388 *esp_sub = 1;
4389 break;
4390 }
4391 if (ops[i].op == OP_MOV && ops[i].operand[0].reg == xAX
4392 && ops[i].operand[1].type == OPT_CONST)
4393 {
4394 for (j = i + 1; j < opcnt; j++)
4395 if (!(ops[j].flags & OPF_DONE))
4396 break;
4397 if (ops[j].op == OP_CALL
4398 && IS(opr_name(&ops[j], 0), "__alloca_probe"))
4399 {
4400 g_stack_fsz += opr_const(&ops[i], 1);
4401 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4402 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4403 i = j + 1;
4404 *esp_sub = 1;
4405 break;
4406 }
4407 }
4408 }
4409
4410 return i;
4411}
4412
4413static void scan_prologue_epilogue(int opcnt, int *stack_align)
4414{
4415 int ecx_push = 0, esp_sub = 0, pusha = 0;
4416 int sandard_epilogue;
4417 int found, ret, len;
4418 int push_fsz = 0;
4419 int i, j, l;
4420
4421 if (g_seh_found == 2) {
4422 eliminate_seh_calls(opcnt);
4423 return;
4424 }
4425 if (g_seh_found) {
4426 eliminate_seh(opcnt);
4427 // ida treats seh as part of sf
4428 g_stack_fsz = g_seh_size;
4429 esp_sub = 1;
4430 }
4431
4432 if (ops[0].op == OP_PUSH && IS(opr_name(&ops[0], 0), "ebp")
4433 && ops[1].op == OP_MOV
4434 && IS(opr_name(&ops[1], 0), "ebp")
4435 && IS(opr_name(&ops[1], 1), "esp"))
4436 {
4437 g_bp_frame = 1;
4438 ops[0].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4439 ops[1].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4440
4441 for (i = 2; i < opcnt; i++)
4442 if (!(ops[i].flags & OPF_DONE))
4443 break;
4444
4445 if (ops[i].op == OP_PUSHA) {
4446 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4447 pusha = 1;
4448 i++;
4449 }
4450
4451 if (ops[i].op == OP_AND && ops[i].operand[0].reg == xSP
4452 && ops[i].operand[1].type == OPT_CONST)
4453 {
4454 l = ops[i].operand[1].val;
4455 j = ffs(l) - 1;
4456 if (j == -1 || (l >> j) != -1)
4457 ferr(&ops[i], "unhandled esp align: %x\n", l);
4458 if (stack_align != NULL)
4459 *stack_align = 1 << j;
4460 ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4461 i++;
4462 }
4463
4464 i = scan_prologue(i, opcnt, &ecx_push, &esp_sub);
4465
4466 found = 0;
4467 do {
4468 for (; i < opcnt; i++)
4469 if (ops[i].flags & OPF_TAIL)
4470 break;
4471 j = i - 1;
4472 if (i == opcnt && (ops[j].flags & OPF_JMP)) {
4473 if (ops[j].bt_i != -1 || ops[j].btj != NULL)
4474 break;
4475 i--;
4476 j--;
4477 }
4478
4479 sandard_epilogue = 0;
4480 if (ops[j].op == OP_POP && IS(opr_name(&ops[j], 0), "ebp"))
4481 {
4482 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4483 // the standard epilogue is sometimes even used without a sf
4484 if (ops[j - 1].op == OP_MOV
4485 && IS(opr_name(&ops[j - 1], 0), "esp")
4486 && IS(opr_name(&ops[j - 1], 1), "ebp"))
4487 sandard_epilogue = 1;
4488 }
4489 else if (ops[j].op == OP_LEAVE)
4490 {
4491 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4492 sandard_epilogue = 1;
4493 }
4494 else if (ops[i].op == OP_CALL && ops[i].pp != NULL
4495 && ops[i].pp->is_noreturn)
4496 {
4497 // on noreturn, msvc sometimes cleans stack, sometimes not
4498 i++;
4499 found = 1;
4500 continue;
4501 }
4502 else if (!(g_ida_func_attr & IDAFA_NORETURN))
4503 ferr(&ops[j], "'pop ebp' expected\n");
4504
4505 if (g_stack_fsz != 0 || sandard_epilogue) {
4506 if (ops[j].op == OP_LEAVE)
4507 j--;
4508 else if (sandard_epilogue) // mov esp, ebp
4509 {
4510 ops[j - 1].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4511 j -= 2;
4512 }
4513 else if (!(g_ida_func_attr & IDAFA_NORETURN))
4514 {
4515 ferr(&ops[j], "esp restore expected\n");
4516 }
4517
4518 if (ecx_push && j >= 0 && ops[j].op == OP_POP
4519 && IS(opr_name(&ops[j], 0), "ecx"))
4520 {
4521 ferr(&ops[j], "unexpected ecx pop\n");
4522 }
4523 }
4524
4525 if (pusha) {
4526 if (ops[j].op == OP_POPA)
4527 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4528 else
4529 ferr(&ops[j], "popa expected\n");
4530 }
4531
4532 found = 1;
4533 i++;
4534 } while (i < opcnt);
4535
4536 if (!found)
4537 ferr(ops, "missing ebp epilogue\n");
4538 return;
4539 }
4540
4541 // non-bp frame
4542 check_simple_sequence(opcnt, &push_fsz);
4543 i = scan_prologue(0, opcnt, &ecx_push, &esp_sub);
4544
4545 found = 0;
4546 if (ecx_push || esp_sub)
4547 {
4548 g_sp_frame = 1;
4549
4550 do {
4551 for (; i < opcnt; i++)
4552 if (ops[i].flags & OPF_TAIL)
4553 break;
4554
4555 j = i - 1;
4556 if (i == opcnt && (ops[j].flags & OPF_JMP)) {
4557 if (ops[j].bt_i != -1 || ops[j].btj != NULL)
4558 break;
4559 i--;
4560 j--;
4561 }
4562 else if (i < opcnt && (ops[i].flags & OPF_ATAIL)) {
4563 // skip arg updates for arg-reuse tailcall
4564 for (; j >= 0; j--) {
4565 if (ops[j].op != OP_MOV)
4566 break;
4567 if (ops[j].operand[0].type == OPT_REGMEM
4568 && strstr(ops[j].operand[0].name, "arg_") != NULL)
4569 continue;
4570 if (ops[j].operand[0].type == OPT_REG)
4571 continue; // assume arg-reg mov
4572 break;
4573 }
4574 }
4575
4576 for (; j >= 0; j--) {
4577 if ((ops[j].flags & (OPF_RMD | OPF_DONE | OPF_NOREGS)) !=
4578 (OPF_RMD | OPF_DONE | OPF_NOREGS))
4579 break;
4580 }
4581
4582 if (ecx_push > 0 && !esp_sub) {
4583 for (l = 0; l < ecx_push && j >= 0; l++) {
4584 if (ops[j].op == OP_POP && IS(opr_name(&ops[j], 0), "ecx"))
4585 /* pop ecx */;
4586 else if (ops[j].op == OP_ADD
4587 && IS(opr_name(&ops[j], 0), "esp")
4588 && ops[j].operand[1].type == OPT_CONST)
4589 {
4590 /* add esp, N */
4591 l += ops[j].operand[1].val / 4 - 1;
4592 }
4593 else
4594 break;
4595
4596 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4597 j--;
4598 }
4599 if (l != ecx_push) {
4600 if (i < opcnt && ops[i].op == OP_CALL
4601 && ops[i].pp != NULL && ops[i].pp->is_noreturn)
4602 {
4603 // noreturn tailcall with no epilogue
4604 i++;
4605 found = 1;
4606 continue;
4607 }
4608 ferr(&ops[j], "epilogue scan failed\n");
4609 }
4610
4611 found = 1;
4612 }
4613
4614 if (esp_sub) {
4615 if (ops[j].op == OP_ADD
4616 && IS(opr_name(&ops[j], 0), "esp")
4617 && ops[j].operand[1].type == OPT_CONST)
4618 {
4619 if (ops[j].operand[1].val < g_stack_fsz)
4620 ferr(&ops[j], "esp adj is too low (need %d)\n", g_stack_fsz);
4621
4622 ops[j].operand[1].val -= g_stack_fsz; // for stack arg scanner
4623 if (ops[j].operand[1].val == 0)
4624 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4625 found = 1;
4626 }
4627 else if (ops[j].op == OP_LEA && ops[j].operand[0].reg == xSP
4628 && ops[j].operand[1].type == OPT_REGMEM
4629 && IS_START(ops[j].operand[1].name, "esp+"))
4630 {
4631 const char *name = ops[j].operand[1].name;
4632 ret = sscanf(name, "esp+%x%n", &l, &len);
4633 ferr_assert(&ops[j], ret == 1 && len == strlen(name));
4634 ferr_assert(&ops[j], l <= g_stack_fsz);
4635 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
4636 found = 1;
4637 }
4638 else if (i < opcnt && ops[i].op == OP_CALL
4639 && ops[i].pp != NULL && ops[i].pp->is_noreturn)
4640 {
4641 // noreturn tailcall with no epilogue
4642 found = 1;
4643 }
4644 else
4645 ferr(&ops[j], "'add esp' expected\n");
4646 }
4647
4648 i++;
4649 } while (i < opcnt);
4650
4651 if (!found)
4652 ferr(ops, "missing esp epilogue\n");
4653 }
4654
4655 if (g_stack_fsz != 0)
4656 // see check_simple_sequence
4657 g_stack_fsz += push_fsz;
4658}
4659
4660// find an instruction that changed opr before i op
4661// *op_i must be set to -1 by the caller
4662// *is_caller is set to 1 if one source is determined to be g_func arg
4663// returns 1 if found, *op_i is then set to origin
4664// returns -1 if multiple origins are found
4665static int resolve_origin(int i, const struct parsed_opr *opr,
4666 int magic, int *op_i, int *is_caller)
4667{
4668 struct label_ref *lr;
4669 int ret = 0;
4670
4671 while (1) {
4672 if (g_labels[i] != NULL) {
4673 lr = &g_label_refs[i];
4674 for (; lr != NULL; lr = lr->next) {
4675 check_i(&ops[i], lr->i);
4676 ret |= resolve_origin(lr->i, opr, magic, op_i, is_caller);
4677 }
4678 if (i > 0 && LAST_OP(i - 1))
4679 return ret;
4680 }
4681
4682 i--;
4683 if (i < 0) {
4684 if (is_caller != NULL)
4685 *is_caller = 1;
4686 return -1;
4687 }
4688
4689 if (ops[i].cc_scratch == magic)
4690 return ret;
4691 ops[i].cc_scratch = magic;
4692
4693 if (!(ops[i].flags & OPF_DATA))
4694 continue;
4695 if (!is_opr_modified(opr, &ops[i]))
4696 continue;
4697
4698 if (*op_i >= 0) {
4699 if (*op_i == i || are_ops_same(&ops[*op_i], &ops[i]))
4700 return ret | 1;
4701
4702 return -1;
4703 }
4704
4705 *op_i = i;
4706 return ret | 1;
4707 }
4708}
4709
4710static int resolve_origin_reg(int i, int reg, int magic, int *op_i,
4711 int *is_caller)
4712{
4713 struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, reg);
4714
4715 *op_i = -1;
4716 if (is_caller != NULL)
4717 *is_caller = 0;
4718 return resolve_origin(i, &opr, magic, op_i, is_caller);
4719}
4720
4721// find an instruction that previously referenced opr
4722// if multiple results are found - fail
4723// *op_i must be set to -1 by the caller
4724// returns 1 if found, *op_i is then set to referencer insn
4725static int resolve_last_ref(int i, const struct parsed_opr *opr,
4726 int magic, int *op_i)
4727{
4728 struct label_ref *lr;
4729 int ret = 0;
4730
4731 while (1) {
4732 if (g_labels[i] != NULL) {
4733 lr = &g_label_refs[i];
4734 for (; lr != NULL; lr = lr->next) {
4735 check_i(&ops[i], lr->i);
4736 ret |= resolve_last_ref(lr->i, opr, magic, op_i);
4737 }
4738 if (i > 0 && LAST_OP(i - 1))
4739 return ret;
4740 }
4741
4742 i--;
4743 if (i < 0)
4744 return -1;
4745
4746 if (ops[i].cc_scratch == magic)
4747 return 0;
4748 ops[i].cc_scratch = magic;
4749
4750 if (!is_opr_referenced(opr, &ops[i]))
4751 continue;
4752
4753 if (*op_i >= 0)
4754 return -1;
4755
4756 *op_i = i;
4757 return 1;
4758 }
4759}
4760
4761// adjust datap of all reachable 'op' insns when moving back
4762// returns 1 if at least 1 op was found
4763// returns -1 if path without an op was found
4764static int adjust_prev_op(int i, enum op_op op, int magic, void *datap)
4765{
4766 struct label_ref *lr;
4767 int ret = 0;
4768
4769 if (ops[i].cc_scratch == magic)
4770 return 0;
4771 ops[i].cc_scratch = magic;
4772
4773 while (1) {
4774 if (g_labels[i] != NULL) {
4775 lr = &g_label_refs[i];
4776 for (; lr != NULL; lr = lr->next) {
4777 check_i(&ops[i], lr->i);
4778 ret |= adjust_prev_op(lr->i, op, magic, datap);
4779 }
4780 if (i > 0 && LAST_OP(i - 1))
4781 return ret;
4782 }
4783
4784 i--;
4785 if (i < 0)
4786 return -1;
4787
4788 if (ops[i].cc_scratch == magic)
4789 return 0;
4790 ops[i].cc_scratch = magic;
4791
4792 if (ops[i].op != op)
4793 continue;
4794
4795 ops[i].datap = datap;
4796 return 1;
4797 }
4798}
4799
4800// find next instruction that reads opr
4801// *op_i must be set to -1 by the caller
4802// on return, *op_i is set to first referencer insn
4803// returns 1 if exactly 1 referencer is found
4804static int find_next_read(int i, int opcnt,
4805 const struct parsed_opr *opr, int magic, int *op_i)
4806{
4807 struct parsed_op *po;
4808 int j, ret = 0;
4809
4810 for (; i < opcnt; i++)
4811 {
4812 if (ops[i].cc_scratch == magic)
4813 return ret;
4814 ops[i].cc_scratch = magic;
4815
4816 po = &ops[i];
4817 if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
4818 if (po->btj != NULL) {
4819 // jumptable
4820 for (j = 0; j < po->btj->count; j++) {
4821 check_i(po, po->btj->d[j].bt_i);
4822 ret |= find_next_read(po->btj->d[j].bt_i, opcnt, opr,
4823 magic, op_i);
4824 }
4825 return ret;
4826 }
4827
4828 if (po->flags & OPF_RMD)
4829 continue;
4830 check_i(po, po->bt_i);
4831 if (po->flags & OPF_CJMP) {
4832 ret |= find_next_read(po->bt_i, opcnt, opr, magic, op_i);
4833 if (ret < 0)
4834 return ret;
4835 }
4836 else
4837 i = po->bt_i - 1;
4838 continue;
4839 }
4840
4841 if (!is_opr_read(opr, po)) {
4842 int full_opr = 1;
4843 if (opr->type == OPT_REG && po->operand[0].type == OPT_REG
4844 && opr->reg == po->operand[0].reg && (po->flags & OPF_DATA))
4845 {
4846 full_opr = po->operand[0].lmod >= opr->lmod;
4847 }
4848 if (is_opr_modified(opr, po) && full_opr) {
4849 // it's overwritten
4850 return ret;
4851 }
4852 if (po->flags & OPF_TAIL)
4853 return ret;
4854 continue;
4855 }
4856
4857 if (*op_i >= 0)
4858 return -1;
4859
4860 *op_i = i;
4861 return 1;
4862 }
4863
4864 return 0;
4865}
4866
4867static int find_next_read_reg(int i, int opcnt, int reg,
4868 enum opr_lenmod lmod, int magic, int *op_i)
4869{
4870 struct parsed_opr opr = OPR_INIT(OPT_REG, lmod, reg);
4871
4872 *op_i = -1;
4873 return find_next_read(i, opcnt, &opr, magic, op_i);
4874}
4875
4876// find next instruction that reads opr
4877// *op_i must be set to -1 by the caller
4878// on return, *op_i is set to first flag user insn
4879// returns 1 if exactly 1 flag user is found
4880static int find_next_flag_use(int i, int opcnt, int magic, int *op_i)
4881{
4882 struct parsed_op *po;
4883 int j, ret = 0;
4884
4885 for (; i < opcnt; i++)
4886 {
4887 if (ops[i].cc_scratch == magic)
4888 return ret;
4889 ops[i].cc_scratch = magic;
4890
4891 po = &ops[i];
4892 if (po->op == OP_CALL)
4893 return -1;
4894 if (po->flags & OPF_JMP) {
4895 if (po->btj != NULL) {
4896 // jumptable
4897 for (j = 0; j < po->btj->count; j++) {
4898 check_i(po, po->btj->d[j].bt_i);
4899 ret |= find_next_flag_use(po->btj->d[j].bt_i, opcnt,
4900 magic, op_i);
4901 }
4902 return ret;
4903 }
4904
4905 if (po->flags & OPF_RMD)
4906 continue;
4907 check_i(po, po->bt_i);
4908 if (po->flags & OPF_CJMP)
4909 goto found;
4910 else
4911 i = po->bt_i - 1;
4912 continue;
4913 }
4914
4915 if (!(po->flags & OPF_CC)) {
4916 if (po->flags & OPF_FLAGS)
4917 // flags changed
4918 return ret;
4919 if (po->flags & OPF_TAIL)
4920 return ret;
4921 continue;
4922 }
4923
4924found:
4925 if (*op_i >= 0)
4926 return -1;
4927
4928 *op_i = i;
4929 return 1;
4930 }
4931
4932 return 0;
4933}
4934
4935static int try_resolve_const(int i, const struct parsed_opr *opr,
4936 int magic, unsigned int *val)
4937{
4938 int s_i = -1;
4939 int ret;
4940
4941 ret = resolve_origin(i, opr, magic, &s_i, NULL);
4942 if (ret == 1) {
4943 i = s_i;
4944 if (ops[i].op != OP_MOV && ops[i].operand[1].type != OPT_CONST)
4945 return -1;
4946
4947 *val = ops[i].operand[1].val;
4948 return 1;
4949 }
4950
4951 return -1;
4952}
4953
4954static int resolve_used_bits(int i, int opcnt, int reg,
4955 int *mask, int *is_z_check)
4956{
4957 struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_WORD, reg);
4958 int j = -1, k = -1;
4959 int ret;
4960
4961 ret = find_next_read(i, opcnt, &opr, i + opcnt * 20, &j);
4962 if (ret != 1)
4963 return -1;
4964
4965 find_next_read(j + 1, opcnt, &opr, i + opcnt * 20 + 1, &k);
4966 if (k != -1) {
4967 fnote(&ops[j], "(first read)\n");
4968 ferr(&ops[k], "TODO: bit resolve: multiple readers\n");
4969 }
4970
4971 if (ops[j].op != OP_TEST || ops[j].operand[1].type != OPT_CONST)
4972 ferr(&ops[j], "TODO: bit resolve: not a const test\n");
4973
4974 ferr_assert(&ops[j], ops[j].operand[0].type == OPT_REG);
4975 ferr_assert(&ops[j], ops[j].operand[0].reg == reg);
4976
4977 *mask = ops[j].operand[1].val;
4978 if (ops[j].operand[0].lmod == OPLM_BYTE
4979 && ops[j].operand[0].name[1] == 'h')
4980 {
4981 *mask <<= 8;
4982 }
4983 ferr_assert(&ops[j], (*mask & ~0xffff) == 0);
4984
4985 *is_z_check = 0;
4986 ret = find_next_flag_use(j + 1, opcnt, i + opcnt * 20 + 2, &k);
4987 if (ret == 1)
4988 *is_z_check = ops[k].pfo == PFO_Z;
4989
4990 return 0;
4991}
4992
4993static const struct parsed_proto *resolve_deref(int i, int magic,
4994 const struct parsed_opr *opr, int level)
4995{
4996 const struct parsed_proto *pp = NULL;
4997 int from_caller = 0;
4998 char s_reg[4];
4999 int offset = 0;
5000 int len = 0;
5001 int j = -1;
5002 int k = -1;
5003 int reg;
5004 int ret;
5005
5006 ret = sscanf(opr->name, "%3s+%x%n", s_reg, &offset, &len);
5007 if (ret != 2 || len != strlen(opr->name)) {
5008 ret = sscanf(opr->name, "%3s%n", s_reg, &len);
5009 if (ret != 1 || len != strlen(opr->name))
5010 return NULL;
5011 }
5012
5013 reg = char_array_i(regs_r32, ARRAY_SIZE(regs_r32), s_reg);
5014 if (reg < 0)
5015 return NULL;
5016
5017 ret = resolve_origin_reg(i, reg, i + magic, &j, NULL);
5018 if (ret != 1)
5019 return NULL;
5020
5021 if (ops[j].op == OP_MOV && ops[j].operand[1].type == OPT_REGMEM
5022 && strlen(ops[j].operand[1].name) == 3
5023 && ops[j].operand[0].lmod == OPLM_DWORD
5024 && ops[j].pp == NULL // no hint
5025 && level == 0)
5026 {
5027 // allow one simple dereference (com/directx)
5028 reg = char_array_i(regs_r32, ARRAY_SIZE(regs_r32),
5029 ops[j].operand[1].name);
5030 if (reg < 0)
5031 return NULL;
5032 ret = resolve_origin_reg(j, reg, j + magic, &k, NULL);
5033 if (ret != 1)
5034 return NULL;
5035 j = k;
5036 }
5037 if (ops[j].op != OP_MOV || ops[j].operand[0].lmod != OPLM_DWORD)
5038 return NULL;
5039
5040 if (ops[j].pp != NULL) {
5041 // type hint in asm
5042 pp = ops[j].pp;
5043 }
5044 else if (ops[j].operand[1].type == OPT_REGMEM) {
5045 pp = try_recover_pp(&ops[j], &ops[j].operand[1], 0, NULL);
5046 if (pp == NULL) {
5047 // maybe structure ptr in structure
5048 pp = resolve_deref(j, magic, &ops[j].operand[1], level + 1);
5049 }
5050 }
5051 else if (ops[j].operand[1].type == OPT_LABEL)
5052 pp = proto_parse(g_fhdr, ops[j].operand[1].name, g_quiet_pp);
5053 else if (ops[j].operand[1].type == OPT_REG) {
5054 // maybe arg reg?
5055 k = -1;
5056 ret = resolve_origin(j, &ops[j].operand[1], i + magic,
5057 &k, &from_caller);
5058 if (ret != 1 && from_caller && k == -1 && g_func_pp != NULL) {
5059 for (k = 0; k < g_func_pp->argc; k++) {
5060 if (g_func_pp->arg[k].reg == NULL)
5061 continue;
5062 if (IS(g_func_pp->arg[k].reg, ops[j].operand[1].name)) {
5063 pp = g_func_pp->arg[k].pp;
5064 break;
5065 }
5066 }
5067 }
5068 }
5069
5070 if (pp == NULL)
5071 return NULL;
5072 if (pp->is_func || pp->is_fptr || !pp->type.is_struct) {
5073 if (offset != 0)
5074 ferr(&ops[j], "expected struct, got '%s %s'\n",
5075 pp->type.name, pp->name);
5076 return NULL;
5077 }
5078
5079 return proto_lookup_struct(g_fhdr, pp->type.name, offset);
5080}
5081
5082static const struct parsed_proto *resolve_func_ptr(int i, int opcnt,
5083 int is_call_op, const struct parsed_opr *opr,
5084 int *pp_i, int *multi_src)
5085{
5086 const struct parsed_proto *pp = NULL;
5087 int search_advice = 0;
5088
5089 if (multi_src != NULL)
5090 *multi_src = 0;
5091 if (pp_i != NULL)
5092 *pp_i = -1;
5093
5094 switch (opr->type) {
5095 case OPT_REGMEM:
5096 // try to resolve struct member calls
5097 pp = resolve_deref(i, i + opcnt * 19, opr, 0);
5098 if (pp != NULL)
5099 break;
5100 // fallthrough
5101 case OPT_LABEL:
5102 case OPT_OFFSET:
5103 pp = try_recover_pp(&ops[i], opr, is_call_op, &search_advice);
5104 if (!search_advice)
5105 break;
5106 // fallthrough
5107 default:
5108 scan_for_call_type(i, opr, i + opcnt * 9, is_call_op,
5109 &pp, pp_i, multi_src);
5110 break;
5111 }
5112
5113 return pp;
5114}
5115
5116static struct parsed_proto *process_call_early(int i, int opcnt,
5117 int *adj_i)
5118{
5119 struct parsed_op *po = &ops[i];
5120 struct parsed_proto *pp;
5121 int multipath = 0;
5122 int adj = 0;
5123 int j, ret;
5124
5125 pp = po->pp;
5126 if (pp == NULL || pp->is_vararg || pp->argc_reg != 0)
5127 // leave for later
5128 return NULL;
5129
5130 // look for and make use of esp adjust
5131 *adj_i = ret = -1;
5132 if (!pp->is_stdcall && pp->argc_stack > 0)
5133 ret = scan_for_esp_adjust(i + 1, opcnt,
5134 pp->argc_stack * 4, &adj, &multipath, 0);
5135 if (ret >= 0) {
5136 if (pp->argc_stack > adj / 4)
5137 return NULL;
5138 if (multipath)
5139 return NULL;
5140 if (ops[ret].op == OP_POP) {
5141 for (j = 1; j < adj / 4; j++) {
5142 if (ops[ret + j].op != OP_POP
5143 || ops[ret + j].operand[0].reg != xCX)
5144 {
5145 return NULL;
5146 }
5147 }
5148 }
5149 }
5150
5151 *adj_i = ret;
5152 return pp;
5153}
5154
5155static struct parsed_proto *process_call(int i, int opcnt)
5156{
5157 struct parsed_op *po = &ops[i];
5158 const struct parsed_proto *pp_c;
5159 struct parsed_proto *pp;
5160 const char *tmpname;
5161 int call_i = -1, ref_i = -1;
5162 int adj = 0, multipath = 0;
5163 int ret, arg;
5164
5165 tmpname = opr_name(po, 0);
5166 pp = po->pp;
5167 if (pp == NULL)
5168 {
5169 // indirect call
5170 pp_c = resolve_func_ptr(i, opcnt, 1, &ops[i].operand[0],
5171 &call_i, &multipath);
5172 if (pp_c != NULL) {
5173 if (!pp_c->is_func && !pp_c->is_fptr)
5174 ferr(po, "call to non-func: %s\n", pp_c->name);
5175 pp = proto_clone(pp_c);
5176 my_assert_not(pp, NULL);
5177 if (multipath)
5178 // not resolved just to single func
5179 pp->is_fptr = 1;
5180
5181 switch (po->operand[0].type) {
5182 case OPT_REG:
5183 // we resolved this call and no longer need the register
5184 po->regmask_src &= ~(1 << po->operand[0].reg);
5185
5186 if (!multipath && i != call_i && ops[call_i].op == OP_MOV
5187 && ops[call_i].operand[1].type == OPT_LABEL)
5188 {
5189 // no other source users?
5190 ret = resolve_last_ref(i, &po->operand[0], i + opcnt * 10,
5191 &ref_i);
5192 if (ret == 1 && call_i == ref_i) {
5193 // and nothing uses it after us?
5194 ref_i = -1;
5195 find_next_read(i + 1, opcnt, &po->operand[0],
5196 i + opcnt * 11, &ref_i);
5197 if (ref_i == -1)
5198 // then also don't need the source mov
5199 ops[call_i].flags |= OPF_RMD | OPF_NOREGS;
5200 }
5201 }
5202 break;
5203 case OPT_REGMEM:
5204 pp->is_fptr = 1;
5205 break;
5206 default:
5207 break;
5208 }
5209 }
5210 if (pp == NULL) {
5211 pp = calloc(1, sizeof(*pp));
5212 my_assert_not(pp, NULL);
5213
5214 pp->is_fptr = 1;
5215 ret = scan_for_esp_adjust(i + 1, opcnt,
5216 -1, &adj, &multipath, 0);
5217 if (ret < 0 || adj < 0) {
5218 if (!g_allow_regfunc)
5219 ferr(po, "non-__cdecl indirect call unhandled yet\n");
5220 pp->is_unresolved = 1;
5221 adj = 0;
5222 }
5223 adj /= 4;
5224 if (adj > ARRAY_SIZE(pp->arg))
5225 ferr(po, "esp adjust too large: %d\n", adj);
5226 pp->ret_type.name = strdup("int");
5227 pp->argc = pp->argc_stack = adj;
5228 for (arg = 0; arg < pp->argc; arg++)
5229 pp->arg[arg].type.name = strdup("int");
5230 }
5231 po->pp = pp;
5232 }
5233
5234 // look for and make use of esp adjust
5235 multipath = 0;
5236 ret = -1;
5237 if (!pp->is_stdcall && pp->argc_stack > 0) {
5238 int adj_expect = pp->is_vararg ? -1 : pp->argc_stack * 4;
5239 ret = scan_for_esp_adjust(i + 1, opcnt,
5240 adj_expect, &adj, &multipath, 0);
5241 }
5242 if (ret >= 0) {
5243 if (pp->is_vararg) {
5244 if (adj / 4 < pp->argc_stack) {
5245 fnote(po, "(this call)\n");
5246 ferr(&ops[ret], "esp adjust is too small: %x < %x\n",
5247 adj, pp->argc_stack * 4);
5248 }
5249 // modify pp to make it have varargs as normal args
5250 arg = pp->argc;
5251 pp->argc += adj / 4 - pp->argc_stack;
5252 for (; arg < pp->argc; arg++) {
5253 pp->arg[arg].type.name = strdup("int");
5254 pp->argc_stack++;
5255 }
5256 if (pp->argc > ARRAY_SIZE(pp->arg))
5257 ferr(po, "too many args for '%s'\n", tmpname);
5258 }
5259 if (pp->argc_stack > adj / 4) {
5260 if (pp->is_noreturn)
5261 // assume no stack adjust was emited
5262 goto out;
5263 fnote(po, "(this call)\n");
5264 ferr(&ops[ret], "stack tracking failed for '%s': %x %x\n",
5265 tmpname, pp->argc_stack * 4, adj);
5266 }
5267
5268 scan_for_esp_adjust(i + 1, opcnt,
5269 pp->argc_stack * 4, &adj, &multipath, 1);
5270 }
5271 else if (pp->is_vararg)
5272 ferr(po, "missing esp_adjust for vararg func '%s'\n",
5273 pp->name);
5274
5275out:
5276 return pp;
5277}
5278
5279static void check_fptr_args(int i, int opcnt, struct parsed_proto *pp)
5280{
5281 struct parsed_opr s_opr = OPR_INIT(OPT_REG, OPLM_DWORD, 0);
5282 const struct parsed_proto *pp_arg, *pp_cmp;
5283 const struct parsed_op *po_a;
5284 const char *s_reg;
5285 int pp_cmp_i;
5286 int arg, reg;
5287 int bad = 0;
5288 int j;
5289
5290 for (arg = 0; arg < pp->argc; arg++) {
5291 pp_cmp = NULL;
5292 pp_cmp_i = -1;
5293
5294 pp_arg = pp->arg[arg].pp;
5295 if (pp_arg == NULL || !pp_arg->is_func)
5296 continue;
5297
5298 s_reg = pp->arg[arg].reg;
5299 if (s_reg != NULL) {
5300 reg = char_array_i(regs_r32, ARRAY_SIZE(regs_r32), s_reg);
5301 ferr_assert(&ops[i], reg >= 0);
5302 s_opr.reg = reg;
5303 scan_for_call_type(i, &s_opr, i + arg + opcnt * 28, 0,
5304 &pp_cmp, &pp_cmp_i, NULL);
5305 if (pp_cmp != NULL && !pp_compatible_func(pp_arg, pp_cmp)) {
5306 bad = 1;
5307 if (pp_cmp_i >= 0)
5308 fnote(&ops[pp_cmp_i], "(referenced here)\n");
5309 }
5310 }
5311 else {
5312 for (j = 0; j < pp->arg[arg].push_ref_cnt; j++) {
5313 po_a = pp->arg[arg].push_refs[j];
5314 if (po_a == NULL || po_a->op != OP_PUSH)
5315 continue;
5316 pp_cmp = resolve_func_ptr(po_a - ops, opcnt, 0,
5317 &po_a->operand[0], &pp_cmp_i, NULL);
5318 if (pp_cmp != NULL && !pp_compatible_func(pp_arg, pp_cmp)) {
5319 bad = 1;
5320 if (pp_cmp_i < 0)
5321 pp_cmp_i = po_a - ops;
5322 if (pp_cmp_i >= 0)
5323 fnote(&ops[pp_cmp_i], "(referenced here)\n");
5324 }
5325 }
5326 }
5327
5328 if (bad)
5329 ferr(&ops[i], "incompatible fptr arg %d\n", arg + 1);
5330 }
5331}
5332
5333static void pp_insert_reg_arg(struct parsed_proto *pp, const char *reg)
5334{
5335 int i;
5336
5337 for (i = 0; i < pp->argc; i++)
5338 if (pp->arg[i].reg == NULL)
5339 break;
5340
5341 if (pp->argc_stack)
5342 memmove(&pp->arg[i + 1], &pp->arg[i],
5343 sizeof(pp->arg[0]) * pp->argc_stack);
5344 memset(&pp->arg[i], 0, sizeof(pp->arg[i]));
5345 pp->arg[i].reg = strdup(reg);
5346 pp->arg[i].type.name = strdup("int");
5347 pp->argc++;
5348 pp->argc_reg++;
5349}
5350
5351static void pp_insert_stack_args(struct parsed_proto *pp, int count)
5352{
5353 int a;
5354
5355 pp->argc += count;
5356 pp->argc_stack += count;
5357
5358 for (a = 0; a < pp->argc; a++)
5359 if (pp->arg[a].type.name == NULL)
5360 pp->arg[a].type.name = strdup("int");
5361}
5362
5363static void pp_add_push_ref(struct parsed_proto *pp,
5364 int arg, struct parsed_op *po)
5365{
5366 pp->arg[arg].push_refs = realloc(pp->arg[arg].push_refs,
5367 (pp->arg[arg].push_ref_cnt + 1)
5368 * sizeof(pp->arg[arg].push_refs[0]));
5369 ferr_assert(po, pp->arg[arg].push_refs != NULL);
5370 pp->arg[arg].push_refs[pp->arg[arg].push_ref_cnt++] = po;
5371}
5372
5373static void mark_float_arg(struct parsed_op *po,
5374 struct parsed_proto *pp, int arg, int *regmask_ffca)
5375{
5376 ferr_assert(po, pp->arg[arg].push_ref_cnt == 0);
5377 pp_add_push_ref(pp, arg, po);
5378
5379 po->p_argnum = arg + 1;
5380 po->flags |= OPF_DONE | OPF_FARGNR | OPF_FARG;
5381 if (regmask_ffca != NULL)
5382 *regmask_ffca |= 1 << arg;
5383}
5384
5385static int check_for_stp(int i, int i_to)
5386{
5387 struct parsed_op *po;
5388
5389 for (; i < i_to; i++) {
5390 po = &ops[i];
5391 if (po->op == OP_FST)
5392 return i;
5393 if (g_labels[i] != NULL || (po->flags & OPF_JMP))
5394 return -1;
5395 if (po->op == OP_CALL || po->op == OP_PUSH || po->op == OP_POP)
5396 return -1;
5397 if (po->op == OP_ADD && po->operand[0].reg == xSP)
5398 return -1;
5399 }
5400
5401 return -1;
5402}
5403
5404static int collect_call_args_no_push(int i, struct parsed_proto *pp,
5405 int *regmask_ffca)
5406{
5407 struct parsed_op *po;
5408 int offset = 0;
5409 int base_arg;
5410 int j, arg;
5411 int ret;
5412
5413 for (base_arg = 0; base_arg < pp->argc; base_arg++)
5414 if (pp->arg[base_arg].reg == NULL)
5415 break;
5416
5417 for (j = i; j > 0; )
5418 {
5419 ferr_assert(&ops[j], g_labels[j] == NULL);
5420 j--;
5421
5422 po = &ops[j];
5423 ferr_assert(po, po->op != OP_PUSH);
5424 if (po->op == OP_FST)
5425 {
5426 if (po->operand[0].type != OPT_REGMEM)
5427 continue;
5428 ret = parse_stack_esp_offset(po, po->operand[0].name, &offset);
5429 if (ret != 0)
5430 continue;
5431 if (offset < 0 || offset >= pp->argc_stack * 4 || (offset & 3)) {
5432 //ferr(po, "offset %d, %d args\n", offset, pp->argc_stack);
5433 continue;
5434 }
5435
5436 arg = base_arg + offset / 4;
5437 mark_float_arg(po, pp, arg, regmask_ffca);
5438 }
5439 else if (po->op == OP_SUB && po->operand[0].reg == xSP
5440 && po->operand[1].type == OPT_CONST)
5441 {
5442 po->flags |= OPF_RMD | OPF_DONE | OPF_FARGNR | OPF_FARG;
5443 break;
5444 }
5445 }
5446
5447 for (arg = base_arg; arg < pp->argc; arg++) {
5448 ferr_assert(&ops[i], pp->arg[arg].reg == NULL);
5449 if (pp->arg[arg].push_ref_cnt != 1)
5450 ferr(&ops[i], "arg %d/%d not found or bad\n", arg, pp->argc);
5451 po = pp->arg[arg].push_refs[0];
5452 if (po->operand[0].lmod == OPLM_QWORD)
5453 arg++;
5454 }
5455
5456 return 0;
5457}
5458
5459static int collect_call_args_early(int i, int opcnt,
5460 struct parsed_proto *pp, int *regmask, int *regmask_ffca)
5461{
5462 struct parsed_op *po;
5463 int arg, ret;
5464 int offset;
5465 int j, k;
5466
5467 for (arg = 0; arg < pp->argc; arg++)
5468 if (pp->arg[arg].reg == NULL)
5469 break;
5470
5471 // first see if it can be easily done
5472 for (j = i; j > 0 && arg < pp->argc; )
5473 {
5474 if (g_labels[j] != NULL)
5475 return -1;
5476 j--;
5477
5478 po = &ops[j];
5479 if (po->op == OP_CALL)
5480 return -1;
5481 else if (po->op == OP_ADD && po->operand[0].reg == xSP)
5482 return -1;
5483 else if (po->op == OP_POP)
5484 return -1;
5485 else if (po->flags & OPF_CJMP)
5486 return -1;
5487 else if (po->op == OP_PUSH) {
5488 if (po->flags & (OPF_FARG|OPF_FARGNR))
5489 return -1;
5490 if (!g_header_mode) {
5491 ret = scan_for_mod(po, j + 1, i, 1);
5492 if (ret >= 0)
5493 return -1;
5494 }
5495
5496 if (pp->arg[arg].type.is_va_list)
5497 return -1;
5498
5499 // next arg
5500 for (arg++; arg < pp->argc; arg++)
5501 if (pp->arg[arg].reg == NULL)
5502 break;
5503 }
5504 else if (po->op == OP_SUB && po->operand[0].reg == xSP
5505 && po->operand[1].type == OPT_CONST)
5506 {
5507 if (po->flags & (OPF_RMD|OPF_DONE))
5508 return -1;
5509 if (po->operand[1].val != pp->argc_stack * 4)
5510 ferr(po, "unexpected esp adjust: %d\n",
5511 po->operand[1].val * 4);
5512 ferr_assert(po, pp->argc - arg == pp->argc_stack);
5513 return collect_call_args_no_push(i, pp, regmask_ffca);
5514 }
5515 }
5516
5517 if (arg < pp->argc)
5518 return -1;
5519
5520 // now do it
5521 for (arg = 0; arg < pp->argc; arg++)
5522 if (pp->arg[arg].reg == NULL)
5523 break;
5524
5525 for (j = i; j > 0 && arg < pp->argc; )
5526 {
5527 j--;
5528
5529 if (ops[j].op == OP_PUSH)
5530 {
5531 int ref_handled = 0;
5532
5533 k = check_for_stp(j + 1, i);
5534 if (k != -1) {
5535 // push ecx; fstp dword ptr [esp]
5536 ret = parse_stack_esp_offset(&ops[k],
5537 ops[k].operand[0].name, &offset);
5538 if (ret == 0 && offset == 0) {
5539 if (!pp->arg[arg].type.is_float)
5540 ferr(&ops[i], "arg %d should be float\n", arg + 1);
5541 mark_float_arg(&ops[k], pp, arg, regmask_ffca);
5542 ref_handled = 1;
5543 }
5544 }
5545
5546 if (!ref_handled) {
5547 ferr_assert(&ops[j], pp->arg[arg].push_ref_cnt == 0);
5548 pp_add_push_ref(pp, arg, &ops[j]);
5549 }
5550
5551 if (regmask != NULL && ops[j].operand[0].type == OPT_REG)
5552 *regmask |= 1 << ops[j].operand[0].reg;
5553
5554 ops[j].flags |= OPF_RMD | OPF_DONE | OPF_FARGNR | OPF_FARG;
5555 ops[j].flags &= ~OPF_RSAVE;
5556
5557 // next arg
5558 for (arg++; arg < pp->argc; arg++)
5559 if (pp->arg[arg].reg == NULL)
5560 break;
5561 }
5562 }
5563
5564 if (!g_header_mode)
5565 check_fptr_args(i, opcnt, pp);
5566
5567 return 0;
5568}
5569
5570// ensure all s_a* numbers match for a given func arg in all branches
5571// returns 1 if any changes were made, 0 if not
5572static int sync_argnum(struct parsed_proto *pp, int arg,
5573 int *argnum, int *arggrp)
5574{
5575 struct parsed_op *po_tmp;
5576 int changed = 0;
5577 int i;
5578
5579 // see if other branches don't have higher argnum
5580 for (i = 0; i < pp->arg[arg].push_ref_cnt; i++) {
5581 po_tmp = pp->arg[arg].push_refs[i];
5582 if (*argnum < po_tmp->p_argnum)
5583 *argnum = po_tmp->p_argnum;
5584 if (*arggrp < po_tmp->p_arggrp)
5585 *arggrp = po_tmp->p_arggrp;
5586 }
5587
5588 // make all argnums consistent
5589 for (i = 0; i < pp->arg[arg].push_ref_cnt; i++) {
5590 po_tmp = pp->arg[arg].push_refs[i];
5591 if (po_tmp->p_argnum == 0)
5592 continue;
5593 if (po_tmp->p_argnum != *argnum || po_tmp->p_arggrp != *arggrp) {
5594 po_tmp->p_argnum = *argnum;
5595 po_tmp->p_arggrp = *arggrp;
5596 changed = 1;
5597 }
5598 }
5599
5600 return changed;
5601}
5602
5603static int collect_call_args_r(struct parsed_op *po, int i,
5604 struct parsed_proto *pp, int *regmask,
5605 int arg, int argnum, int magic,
5606 int skip, int need_op_saving, int may_reuse)
5607{
5608 struct parsed_proto *pp_tmp;
5609 struct label_ref *lr;
5610 int need_to_save_current;
5611 int arg_grp_current = 0;
5612 int save_args_seen = 0;
5613 int dummy = 0;
5614 int ret = 0;
5615 int reg;
5616 char buf[32];
5617 int j, k;
5618
5619 if (i < 0) {
5620 ferr(po, "dead label encountered\n");
5621 return -1;
5622 }
5623
5624 for (; arg < pp->argc; arg++, argnum++)
5625 if (pp->arg[arg].reg == NULL)
5626 break;
5627 magic = (magic & 0xffffff) | (arg << 24);
5628
5629 for (j = i; j >= 0 && (arg < pp->argc || pp->is_unresolved); )
5630 {
5631 if (((ops[j].cc_scratch ^ magic) & 0xffffff) == 0) {
5632 if (ops[j].cc_scratch != magic) {
5633 ferr(&ops[j], "arg collect hit same path with diff args for %s\n",
5634 pp->name);
5635 return -1;
5636 }
5637 // ok: have already been here
5638 return 0;
5639 }
5640 ops[j].cc_scratch = magic;
5641
5642 if (g_labels[j] != NULL && g_label_refs[j].i != -1) {
5643 lr = &g_label_refs[j];
5644 if (lr->next != NULL)
5645 need_op_saving = 1;
5646 for (; lr->next; lr = lr->next) {
5647 check_i(&ops[j], lr->i);
5648 if ((ops[lr->i].flags & (OPF_JMP|OPF_CJMP)) != OPF_JMP)
5649 may_reuse = 1;
5650 ret = collect_call_args_r(po, lr->i, pp, regmask,
5651 arg, argnum, magic, skip, need_op_saving, may_reuse);
5652 if (ret < 0)
5653 return ret;
5654 }
5655
5656 check_i(&ops[j], lr->i);
5657 if ((ops[lr->i].flags & (OPF_JMP|OPF_CJMP)) != OPF_JMP)
5658 may_reuse = 1;
5659 if (j > 0 && LAST_OP(j - 1)) {
5660 // follow last branch in reverse
5661 j = lr->i;
5662 continue;
5663 }
5664 need_op_saving = 1;
5665 ret = collect_call_args_r(po, lr->i, pp, regmask,
5666 arg, argnum, magic, skip, need_op_saving, may_reuse);
5667 if (ret < 0)
5668 return ret;
5669 }
5670 j--;
5671
5672 if (ops[j].op == OP_CALL)
5673 {
5674 if (pp->is_unresolved)
5675 break;
5676
5677 pp_tmp = ops[j].pp;
5678 if (pp_tmp == NULL)
5679 ferr(po, "arg collect %d/%d hit unparsed call '%s'\n",
5680 arg, pp->argc, ops[j].operand[0].name);
5681 if (may_reuse && pp_tmp->argc_stack > 0)
5682 ferr(po, "arg collect %d/%d hit '%s' with %d stack args\n",
5683 arg, pp->argc, opr_name(&ops[j], 0), pp_tmp->argc_stack);
5684 if (!pp_tmp->is_unresolved)
5685 skip = pp_tmp->argc_stack;
5686 }
5687 // esp adjust of 0 means we collected it before
5688 else if (ops[j].op == OP_ADD && ops[j].operand[0].reg == xSP
5689 && (ops[j].operand[1].type != OPT_CONST
5690 || ops[j].operand[1].val != 0))
5691 {
5692 if (pp->is_unresolved)
5693 break;
5694
5695 fnote(po, "(this call)\n");
5696 ferr(&ops[j], "arg collect %d/%d hit esp adjust of %d\n",
5697 arg, pp->argc, ops[j].operand[1].val);
5698 }
5699 else if (ops[j].op == OP_POP && !(ops[j].flags & OPF_DONE))
5700 {
5701 if (pp->is_unresolved)
5702 break;
5703
5704 fnote(po, "(this call)\n");
5705 ferr(&ops[j], "arg collect %d/%d hit pop\n", arg, pp->argc);
5706 }
5707 else if (ops[j].flags & OPF_CJMP)
5708 {
5709 if (pp->is_unresolved)
5710 break;
5711
5712 may_reuse = 1;
5713 }
5714 else if (ops[j].op == OP_PUSH && skip > 0) {
5715 // XXX: might want to rm OPF_FARGNR and only use this
5716 skip--;
5717 }
5718 else if (ops[j].op == OP_PUSH
5719 && !(ops[j].flags & (OPF_FARGNR|OPF_DONE)))
5720 {
5721 if (pp->is_unresolved && (ops[j].flags & OPF_RMD))
5722 break;
5723
5724 pp_add_push_ref(pp, arg, &ops[j]);
5725
5726 sync_argnum(pp, arg, &argnum, &dummy);
5727
5728 need_to_save_current = 0;
5729 reg = -1;
5730 if (ops[j].operand[0].type == OPT_REG)
5731 reg = ops[j].operand[0].reg;
5732
5733 if (!need_op_saving) {
5734 ret = scan_for_mod(&ops[j], j + 1, i, 1);
5735 need_to_save_current = (ret >= 0);
5736 }
5737 if (need_op_saving || need_to_save_current) {
5738 // mark this arg as one that needs operand saving
5739 pp->arg[arg].is_saved = 1;
5740
5741 if (save_args_seen & (1 << (argnum - 1))) {
5742 save_args_seen = 0;
5743 arg_grp_current++;
5744 if (arg_grp_current >= MAX_ARG_GRP)
5745 ferr(&ops[j], "out of arg groups (arg%d), f %s\n",
5746 argnum, pp->name);
5747 }
5748 }
5749 else if (ops[j].p_argnum == 0)
5750 ops[j].flags |= OPF_RMD;
5751
5752 // some PUSHes are reused by different calls on other branches,
5753 // but that can't happen if we didn't branch, so they
5754 // can be removed from future searches (handles nested calls)
5755 if (!may_reuse)
5756 ops[j].flags |= OPF_FARGNR;
5757
5758 ops[j].flags |= OPF_FARG;
5759 ops[j].flags &= ~OPF_RSAVE;
5760
5761 // check for __VALIST
5762 if (!pp->is_unresolved && g_func_pp != NULL
5763 && pp->arg[arg].type.is_va_list)
5764 {
5765 k = -1;
5766 ret = resolve_origin(j, &ops[j].operand[0],
5767 magic + 1, &k, NULL);
5768 if (ret == 1 && k >= 0)
5769 {
5770 if (ops[k].op == OP_LEA) {
5771 if (!g_func_pp->is_vararg)
5772 ferr(&ops[k], "lea <arg> used, but %s is not vararg?\n",
5773 g_func_pp->name);
5774
5775 snprintf(buf, sizeof(buf), "arg_%X",
5776 g_func_pp->argc_stack * 4);
5777 if (strstr(ops[k].operand[1].name, buf)
5778 || strstr(ops[k].operand[1].name, "arglist"))
5779 {
5780 ops[k].flags |= OPF_RMD | OPF_NOREGS | OPF_DONE;
5781 ops[j].flags |= OPF_RMD | OPF_NOREGS | OPF_VAPUSH;
5782 pp->arg[arg].is_saved = 0;
5783 reg = -1;
5784 }
5785 else
5786 ferr(&ops[k], "va_list arg detection failed\n");
5787 }
5788 // check for va_list from g_func_pp arg too
5789 else if (ops[k].op == OP_MOV
5790 && is_stack_access(&ops[k], &ops[k].operand[1]))
5791 {
5792 ret = stack_frame_access(&ops[k], &ops[k].operand[1],
5793 buf, sizeof(buf), ops[k].operand[1].name, "", 1, 0);
5794 if (ret >= 0) {
5795 ops[k].flags |= OPF_RMD | OPF_DONE;
5796 ops[j].flags |= OPF_RMD;
5797 ops[j].p_argpass = ret + 1;
5798 pp->arg[arg].is_saved = 0;
5799 reg = -1;
5800 }
5801 }
5802 }
5803 }
5804
5805 if (pp->arg[arg].is_saved) {
5806 ops[j].flags &= ~OPF_RMD;
5807 ops[j].p_argnum = argnum;
5808 ops[j].p_arggrp = arg_grp_current;
5809 }
5810
5811 // tracking reg usage
5812 if (reg >= 0)
5813 *regmask |= 1 << reg;
5814
5815 arg++;
5816 argnum++;
5817 if (!pp->is_unresolved) {
5818 // next arg
5819 for (; arg < pp->argc; arg++, argnum++)
5820 if (pp->arg[arg].reg == NULL)
5821 break;
5822 }
5823 magic = (magic & 0xffffff) | (arg << 24);
5824 }
5825
5826 if (ops[j].p_arggrp > arg_grp_current) {
5827 save_args_seen = 0;
5828 arg_grp_current = ops[j].p_arggrp;
5829 }
5830 if (ops[j].p_argnum > 0)
5831 save_args_seen |= 1 << (ops[j].p_argnum - 1);
5832 }
5833
5834 if (arg < pp->argc) {
5835 ferr(po, "arg collect failed for '%s': %d/%d\n",
5836 pp->name, arg, pp->argc);
5837 return -1;
5838 }
5839
5840 return arg;
5841}
5842
5843static int collect_call_args(struct parsed_op *po, int i, int opcnt,
5844 struct parsed_proto *pp, int *regmask, int magic)
5845{
5846 int ret;
5847
5848 ret = collect_call_args_r(po, i, pp, regmask, 0, 1, magic,
5849 0, 0, 0);
5850 if (ret < 0)
5851 return ret;
5852
5853 if (pp->is_unresolved)
5854 pp_insert_stack_args(pp, ret);
5855
5856 // note: p_argnum, p_arggrp will be propagated in a later pass,
5857 // look for sync_argnum() (p_arggrp is for cases when mixed pushes
5858 // for multiple funcs are going on)
5859
5860 if (!g_header_mode)
5861 check_fptr_args(i, opcnt, pp);
5862
5863 return ret;
5864}
5865
5866static void reg_use_pass(int i, int opcnt, unsigned char *cbits,
5867 int regmask_now, int *regmask,
5868 int regmask_save_now, int *regmask_save,
5869 int *regmask_init, int regmask_arg)
5870{
5871 struct parsed_op *po;
5872 int already_saved;
5873 int regmask_new;
5874 int regmask_op;
5875 int flags_set;
5876 int ret, reg;
5877 int j;
5878
5879 for (; i < opcnt; i++)
5880 {
5881 po = &ops[i];
5882 if (cbits[i >> 3] & (1 << (i & 7)))
5883 return;
5884 cbits[i >> 3] |= (1 << (i & 7));
5885
5886 if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
5887 if (po->flags & (OPF_RMD|OPF_DONE))
5888 continue;
5889 if (po->btj != NULL) {
5890 for (j = 0; j < po->btj->count; j++) {
5891 check_i(po, po->btj->d[j].bt_i);
5892 reg_use_pass(po->btj->d[j].bt_i, opcnt, cbits,
5893 regmask_now, regmask, regmask_save_now, regmask_save,
5894 regmask_init, regmask_arg);
5895 }
5896 return;
5897 }
5898
5899 check_i(po, po->bt_i);
5900 if (po->flags & OPF_CJMP)
5901 reg_use_pass(po->bt_i, opcnt, cbits,
5902 regmask_now, regmask, regmask_save_now, regmask_save,
5903 regmask_init, regmask_arg);
5904 else
5905 i = po->bt_i - 1;
5906 continue;
5907 }
5908
5909 if (po->op == OP_PUSH && !(po->flags & (OPF_FARG|OPF_DONE))
5910 && !g_func_pp->is_userstack
5911 && po->operand[0].type == OPT_REG)
5912 {
5913 int save_level = 0;
5914
5915 reg = po->operand[0].reg;
5916 ferr_assert(po, reg >= 0);
5917
5918 already_saved = 0;
5919 flags_set = OPF_RSAVE | OPF_RMD | OPF_DONE;
5920 if (regmask_now & (1 << reg)) {
5921 already_saved = regmask_save_now & (1 << reg);
5922 flags_set = OPF_RSAVE | OPF_DONE;
5923 save_level++;
5924 }
5925
5926 ret = scan_for_pop(i + 1, opcnt, i + opcnt * 3,
5927 reg, 0, 0, save_level, 0);
5928 if (ret == 1) {
5929 scan_for_pop(i + 1, opcnt, i + opcnt * 4,
5930 reg, 0, 0, save_level, flags_set);
5931 }
5932 else {
5933 ret = scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg, 0);
5934 if (ret == 1) {
5935 scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg,
5936 flags_set);
5937 }
5938 }
5939 if (ret == 1) {
5940 ferr_assert(po, !already_saved);
5941 po->flags |= flags_set;
5942
5943 if (regmask_now & (1 << reg)) {
5944 regmask_save_now |= (1 << reg);
5945 *regmask_save |= regmask_save_now;
5946 }
5947 continue;
5948 }
5949 }
5950 else if (po->op == OP_POP && (po->flags & OPF_RSAVE)) {
5951 reg = po->operand[0].reg;
5952 ferr_assert(po, reg >= 0);
5953
5954 if (regmask_save_now & (1 << reg))
5955 regmask_save_now &= ~(1 << reg);
5956 else
5957 regmask_now &= ~(1 << reg);
5958 continue;
5959 }
5960 else if (po->op == OP_CALL) {
5961 if ((po->regmask_dst & (1 << xAX))
5962 && !(po->regmask_dst & (1 << xDX)))
5963 {
5964 if (po->flags & OPF_TAIL)
5965 // don't need eax, will do "return f();" or "f(); return;"
5966 po->regmask_dst &= ~(1 << xAX);
5967 else {
5968 find_next_read_reg(i + 1, opcnt, xAX, OPLM_DWORD,
5969 i + opcnt * 17, &j);
5970 if (j == -1)
5971 // not used
5972 po->regmask_dst &= ~(1 << xAX);
5973 }
5974 }
5975
5976 // not "full stack" mode and have something in stack
5977 if (!(regmask_now & mxST7_2) && (regmask_now & mxST1_0))
5978 ferr(po, "float stack is not empty on func call\n");
5979 }
5980
5981 if (po->flags & OPF_NOREGS)
5982 continue;
5983
5984 // if incomplete register is used, clear it on init to avoid
5985 // later use of uninitialized upper part in some situations
5986 if ((po->flags & OPF_DATA) && po->operand[0].type == OPT_REG
5987 && po->operand[0].lmod != OPLM_DWORD)
5988 {
5989 reg = po->operand[0].reg;
5990 ferr_assert(po, reg >= 0);
5991
5992 if (!(regmask_now & (1 << reg)))
5993 *regmask_init |= 1 << reg;
5994 }
5995
5996 regmask_op = po->regmask_src | po->regmask_dst;
5997
5998 regmask_new = po->regmask_src & ~regmask_now & ~regmask_arg;
5999 regmask_new &= ~(1 << xSP);
6000 if (g_bp_frame && !(po->flags & OPF_EBP_S))
6001 regmask_new &= ~(1 << xBP);
6002
6003 if (regmask_new != 0)
6004 fnote(po, "uninitialized reg mask: %x\n", regmask_new);
6005
6006 if (regmask_op & (1 << xBP)) {
6007 if (g_bp_frame && !(po->flags & OPF_EBP_S)) {
6008 if (po->regmask_dst & (1 << xBP))
6009 // compiler decided to drop bp frame and use ebp as scratch
6010 scan_fwd_set_flags(i + 1, opcnt, i + opcnt * 5, OPF_EBP_S);
6011 else
6012 regmask_op &= ~(1 << xBP);
6013 }
6014 }
6015
6016 if (po->flags & OPF_FPUSH) {
6017 if (regmask_now & mxST1)
6018 regmask_now |= mxSTa; // switch to "full stack" mode
6019 if (regmask_now & mxSTa)
6020 po->flags |= OPF_FSHIFT;
6021 if (!(regmask_now & mxST7_2)) {
6022 regmask_now =
6023 (regmask_now & ~mxST1_0) | ((regmask_now & mxST0) << 1);
6024 }
6025 }
6026
6027 regmask_now |= regmask_op;
6028 *regmask |= regmask_now;
6029
6030 // released regs
6031 if (po->flags & OPF_FPOPP) {
6032 if ((regmask_now & mxSTa) == 0)
6033 ferr(po, "float pop on empty stack?\n");
6034 if (regmask_now & mxST7_2)
6035 po->flags |= OPF_FSHIFT;
6036 if (!(regmask_now & mxST7_2))
6037 regmask_now &= ~mxST1_0;
6038 }
6039 else if (po->flags & OPF_FPOP) {
6040 if ((regmask_now & mxSTa) == 0)
6041 ferr(po, "float pop on empty stack?\n");
6042 if (regmask_now & (mxST7_2 | mxST1))
6043 po->flags |= OPF_FSHIFT;
6044 if (!(regmask_now & mxST7_2)) {
6045 regmask_now =
6046 (regmask_now & ~mxST1_0) | ((regmask_now & mxST1) >> 1);
6047 }
6048 }
6049
6050 if (po->flags & OPF_TAIL) {
6051 if (!(regmask_now & mxST7_2)) {
6052 if (get_pp_arg_regmask_dst(g_func_pp) & mxST0) {
6053 if (!(regmask_now & mxST0))
6054 ferr(po, "no st0 on float return, mask: %x\n",
6055 regmask_now);
6056 }
6057 else if (regmask_now & mxST1_0)
6058 ferr(po, "float regs on tail: %x\n", regmask_now);
6059 }
6060
6061 // there is support for "conditional tailcall", sort of
6062 if (!(po->flags & OPF_CC))
6063 return;
6064 }
6065 }
6066}
6067
6068static void output_std_flag_z(FILE *fout, struct parsed_op *po,
6069 int *pfomask, const char *dst_opr_text)
6070{
6071 if (*pfomask & (1 << PFO_Z)) {
6072 fprintf(fout, "\n cond_z = (%s%s == 0);",
6073 lmod_cast_u(po, po->operand[0].lmod), dst_opr_text);
6074 *pfomask &= ~(1 << PFO_Z);
6075 }
6076}
6077
6078static void output_std_flag_s(FILE *fout, struct parsed_op *po,
6079 int *pfomask, const char *dst_opr_text)
6080{
6081 if (*pfomask & (1 << PFO_S)) {
6082 fprintf(fout, "\n cond_s = (%s%s < 0);",
6083 lmod_cast_s(po, po->operand[0].lmod), dst_opr_text);
6084 *pfomask &= ~(1 << PFO_S);
6085 }
6086}
6087
6088static void output_std_flags(FILE *fout, struct parsed_op *po,
6089 int *pfomask, const char *dst_opr_text)
6090{
6091 output_std_flag_z(fout, po, pfomask, dst_opr_text);
6092 output_std_flag_s(fout, po, pfomask, dst_opr_text);
6093}
6094
6095enum {
6096 OPP_FORCE_NORETURN = (1 << 0),
6097 OPP_SIMPLE_ARGS = (1 << 1),
6098 OPP_ALIGN = (1 << 2),
6099};
6100
6101static void output_pp_attrs(FILE *fout, const struct parsed_proto *pp,
6102 int flags)
6103{
6104 const char *cconv = "";
6105
6106 if (pp->is_fastcall)
6107 cconv = "__fastcall ";
6108 else if (pp->is_stdcall && pp->argc_reg == 0)
6109 cconv = "__stdcall ";
6110
6111 fprintf(fout, (flags & OPP_ALIGN) ? "%-16s" : "%s", cconv);
6112
6113 if (pp->is_noreturn || (flags & OPP_FORCE_NORETURN))
6114 fprintf(fout, "noreturn ");
6115}
6116
6117static void output_pp(FILE *fout, const struct parsed_proto *pp,
6118 int flags)
6119{
6120 int i;
6121
6122 fprintf(fout, (flags & OPP_ALIGN) ? "%-5s" : "%s ",
6123 pp->ret_type.name);
6124 if (pp->is_fptr)
6125 fprintf(fout, "(");
6126 output_pp_attrs(fout, pp, flags);
6127 if (pp->is_fptr)
6128 fprintf(fout, "*");
6129 fprintf(fout, "%s", pp->name);
6130 if (pp->is_fptr)
6131 fprintf(fout, ")");
6132
6133 fprintf(fout, "(");
6134 for (i = 0; i < pp->argc; i++) {
6135 if (i > 0)
6136 fprintf(fout, ", ");
6137 if (pp->arg[i].pp != NULL && pp->arg[i].pp->is_func
6138 && !(flags & OPP_SIMPLE_ARGS))
6139 {
6140 // func pointer
6141 output_pp(fout, pp->arg[i].pp, 0);
6142 }
6143 else if (pp->arg[i].type.is_retreg) {
6144 fprintf(fout, "u32 *r_%s", pp->arg[i].reg);
6145 }
6146 else {
6147 fprintf(fout, "%s", pp->arg[i].type.name);
6148 if (!pp->is_fptr)
6149 fprintf(fout, " a%d", i + 1);
6150 }
6151
6152 if (pp->arg[i].type.is_64bit)
6153 i++;
6154 }
6155 if (pp->is_vararg) {
6156 if (i > 0)
6157 fprintf(fout, ", ");
6158 fprintf(fout, "...");
6159 }
6160 fprintf(fout, ")");
6161}
6162
6163static char *saved_arg_name(char *buf, size_t buf_size, int grp, int num)
6164{
6165 char buf1[16];
6166
6167 buf1[0] = 0;
6168 if (grp > 0)
6169 snprintf(buf1, sizeof(buf1), "%d", grp);
6170 snprintf(buf, buf_size, "s%s_a%d", buf1, num);
6171
6172 return buf;
6173}
6174
6175static void gen_x_cleanup(int opcnt);
6176
6177static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt)
6178{
6179 struct parsed_op *po, *delayed_flag_op = NULL, *tmp_op;
6180 struct parsed_opr *last_arith_dst = NULL;
6181 char buf1[256], buf2[256], buf3[256], cast[64];
6182 struct parsed_proto *pp, *pp_tmp;
6183 struct parsed_data *pd;
6184 int save_arg_vars[MAX_ARG_GRP] = { 0, };
6185 unsigned char cbits[MAX_OPS / 8];
6186 const char *float_type;
6187 const char *float_st0;
6188 const char *float_st1;
6189 int need_float_stack = 0;
6190 int need_float_sw = 0; // status word
6191 int need_tmp_var = 0;
6192 int need_tmp64 = 0;
6193 int cond_vars = 0;
6194 int had_decl = 0;
6195 int label_pending = 0;
6196 int need_double = 0;
6197 int stack_align = 0;
6198 int stack_fsz_adj = 0;
6199 int lock_handled = 0;
6200 int regmask_save = 0; // used regs saved/restored in this func
6201 int regmask_arg; // regs from this function args (fastcall, etc)
6202 int regmask_ret; // regs needed on ret
6203 int regmask_now; // temp
6204 int regmask_init = 0; // regs that need zero initialization
6205 int regmask_pp = 0; // regs used in complex push-pop graph
6206 int regmask_ffca = 0; // float function call args
6207 int regmask = 0; // used regs
6208 int pfomask = 0;
6209 int found = 0;
6210 int dead_dst;
6211 int no_output;
6212 int i, j, l;
6213 int arg;
6214 int reg;
6215 int ret;
6216
6217 g_bp_frame = g_sp_frame = g_stack_fsz = 0;
6218 g_stack_frame_used = 0;
6219 g_seh_size = 0;
6220 if (g_sct_func_attr & SCTFA_CLEAR_REGS)
6221 regmask_init = g_regmask_init;
6222
6223 g_func_pp = proto_parse(fhdr, funcn, 0);
6224 if (g_func_pp == NULL)
6225 ferr(ops, "proto_parse failed for '%s'\n", funcn);
6226
6227 regmask_arg = get_pp_arg_regmask_src(g_func_pp);
6228 regmask_ret = get_pp_arg_regmask_dst(g_func_pp);
6229
6230 // pass1:
6231 // - resolve all branches
6232 // - parse calls with labels
6233 resolve_branches_parse_calls(opcnt);
6234
6235 // pass2:
6236 // - handle ebp/esp frame, remove ops related to it
6237 scan_prologue_epilogue(opcnt, &stack_align);
6238
6239 // handle a case where sf size is unalignment, but is
6240 // placed in a way that elements are still aligned
6241 if (g_stack_fsz & 4) {
6242 for (i = 0; i < g_eqcnt; i++) {
6243 if (g_eqs[i].lmod != OPLM_QWORD)
6244 continue;
6245 if (!(g_eqs[i].offset & 4)) {
6246 g_stack_fsz += 4;
6247 stack_fsz_adj = 4;
6248 }
6249 break;
6250 }
6251 }
6252
6253 // pass3:
6254 // - remove dead labels
6255 // - set regs needed at ret
6256 for (i = 0; i < opcnt; i++)
6257 {
6258 if (g_labels[i] != NULL && g_label_refs[i].i == -1) {
6259 free(g_labels[i]);
6260 g_labels[i] = NULL;
6261 }
6262
6263 if (ops[i].op == OP_RET)
6264 ops[i].regmask_src |= regmask_ret;
6265 }
6266
6267 // pass4:
6268 // - process trivial calls
6269 for (i = 0; i < opcnt; i++)
6270 {
6271 po = &ops[i];
6272 if (po->flags & (OPF_RMD|OPF_DONE))
6273 continue;
6274
6275 if (po->op == OP_CALL)
6276 {
6277 pp = process_call_early(i, opcnt, &j);
6278 if (pp != NULL) {
6279 if (!(po->flags & OPF_ATAIL)) {
6280 // since we know the args, try to collect them
6281 ret = collect_call_args_early(i, opcnt, pp,
6282 &regmask, &regmask_ffca);
6283 if (ret != 0)
6284 pp = NULL;
6285 }
6286 }
6287
6288 if (pp != NULL) {
6289 if (j >= 0) {
6290 // commit esp adjust
6291 if (ops[j].op != OP_POP)
6292 patch_esp_adjust(&ops[j], pp->argc_stack * 4);
6293 else {
6294 for (l = 0; l < pp->argc_stack; l++)
6295 ops[j + l].flags |= OPF_DONE | OPF_RMD | OPF_NOREGS;
6296 }
6297 }
6298
6299 if (strstr(pp->ret_type.name, "int64"))
6300 need_tmp64 = 1;
6301
6302 po->flags |= OPF_DONE;
6303 }
6304 }
6305 }
6306
6307 // pass5:
6308 // - process calls, stage 2
6309 // - handle some push/pop pairs
6310 // - scan for STD/CLD, propagate DF
6311 // - try to resolve needed x87 status word bits
6312 for (i = 0; i < opcnt; i++)
6313 {
6314 int mask, z_check;
6315
6316 po = &ops[i];
6317 if (po->flags & OPF_RMD)
6318 continue;
6319
6320 if (po->op == OP_CALL)
6321 {
6322 if (!(po->flags & OPF_DONE)) {
6323 pp = process_call(i, opcnt);
6324
6325 if (!pp->is_unresolved && !(po->flags & OPF_ATAIL)) {
6326 // since we know the args, collect them
6327 collect_call_args(po, i, opcnt, pp, &regmask, i + opcnt * 2);
6328 }
6329 // for unresolved, collect after other passes
6330 }
6331
6332 pp = po->pp;
6333 ferr_assert(po, pp != NULL);
6334
6335 po->regmask_src |= get_pp_arg_regmask_src(pp);
6336 po->regmask_dst |= get_pp_arg_regmask_dst(pp);
6337
6338 if (po->regmask_dst & mxST0)
6339 po->flags |= OPF_FPUSH;
6340
6341 if (strstr(pp->ret_type.name, "int64"))
6342 need_tmp64 = 1;
6343
6344 continue;
6345 }
6346
6347 if (po->flags & OPF_DONE)
6348 continue;
6349
6350 switch (po->op) {
6351 case OP_PUSH:
6352 if (!(po->flags & OPF_FARG) && !(po->flags & OPF_RSAVE)
6353 && po->operand[0].type == OPT_CONST)
6354 {
6355 scan_for_pop_const(i, opcnt, i + opcnt * 12);
6356 }
6357 break;
6358
6359 case OP_POP:
6360 scan_pushes_for_pop(i, opcnt, &regmask_pp);
6361 break;
6362
6363 case OP_STD:
6364 po->flags |= OPF_DF | OPF_RMD | OPF_DONE;
6365 scan_propagate_df(i + 1, opcnt);
6366 break;
6367
6368 case OP_FNSTSW:
6369 need_float_sw = 1;
6370 if (po->operand[0].type != OPT_REG || po->operand[0].reg != xAX)
6371 ferr(po, "TODO: fnstsw to mem\n");
6372 ret = resolve_used_bits(i + 1, opcnt, xAX, &mask, &z_check);
6373 if (ret != 0)
6374 ferr(po, "fnstsw resolve failed\n");
6375 ret = adjust_prev_op(i, OP_FCOM, i + opcnt * 21,
6376 (void *)(long)(mask | (z_check << 16)));
6377 if (ret != 1)
6378 ferr(po, "failed to find fcom: %d\n", ret);
6379 break;
6380
6381 default:
6382 break;
6383 }
6384 }
6385
6386 // pass6:
6387 // - find POPs for PUSHes, rm both
6388 // - scan for all used registers
6389 memset(cbits, 0, sizeof(cbits));
6390 reg_use_pass(0, opcnt, cbits, regmask_init, &regmask,
6391 0, &regmask_save, &regmask_init, regmask_arg);
6392
6393 need_float_stack = !!(regmask & mxST7_2);
6394
6395 // pass7:
6396 // - find flag set ops for their users
6397 // - do unresolved calls
6398 // - declare indirect functions
6399 // - other op specific processing
6400 for (i = 0; i < opcnt; i++)
6401 {
6402 po = &ops[i];
6403 if (po->flags & (OPF_RMD|OPF_DONE))
6404 continue;
6405
6406 if (po->flags & OPF_CC)
6407 {
6408 int setters[16], cnt = 0, branched = 0;
6409
6410 ret = scan_for_flag_set(i, opcnt, i + opcnt * 6,
6411 &branched, setters, &cnt);
6412 if (ret < 0 || cnt <= 0)
6413 ferr(po, "unable to trace flag setter(s)\n");
6414 if (cnt > ARRAY_SIZE(setters))
6415 ferr(po, "too many flag setters\n");
6416
6417 for (j = 0; j < cnt; j++)
6418 {
6419 tmp_op = &ops[setters[j]]; // flag setter
6420 pfomask = 0;
6421
6422 // to get nicer code, we try to delay test and cmp;
6423 // if we can't because of operand modification, or if we
6424 // have arith op, or branch, make it calculate flags explicitly
6425 if (tmp_op->op == OP_TEST || tmp_op->op == OP_CMP)
6426 {
6427 if (branched || scan_for_mod(tmp_op, setters[j] + 1, i, 0) >= 0)
6428 pfomask = 1 << po->pfo;
6429 }
6430 else if (tmp_op->op == OP_CMPS || tmp_op->op == OP_SCAS) {
6431 pfomask = 1 << po->pfo;
6432 }
6433 else {
6434 // see if we'll be able to handle based on op result
6435 if ((tmp_op->op != OP_AND && tmp_op->op != OP_OR
6436 && po->pfo != PFO_Z && po->pfo != PFO_S
6437 && po->pfo != PFO_P)
6438 || branched
6439 || scan_for_mod_opr0(tmp_op, setters[j] + 1, i) >= 0)
6440 {
6441 pfomask = 1 << po->pfo;
6442 }
6443
6444 if (tmp_op->op == OP_ADD && po->pfo == PFO_C) {
6445 propagate_lmod(tmp_op, &tmp_op->operand[0],
6446 &tmp_op->operand[1]);
6447 if (tmp_op->operand[0].lmod == OPLM_DWORD)
6448 need_tmp64 = 1;
6449 }
6450 }
6451 if (pfomask) {
6452 tmp_op->pfomask |= pfomask;
6453 cond_vars |= pfomask;
6454 }
6455 // note: may overwrite, currently not a problem
6456 po->datap = tmp_op;
6457 }
6458
6459 if (po->op == OP_RCL || po->op == OP_RCR
6460 || po->op == OP_ADC || po->op == OP_SBB)
6461 cond_vars |= 1 << PFO_C;
6462 }
6463
6464 switch (po->op) {
6465 case OP_CMPS:
6466 case OP_SCAS:
6467 cond_vars |= 1 << PFO_Z;
6468 break;
6469
6470 case OP_MUL:
6471 if (po->operand[0].lmod == OPLM_DWORD)
6472 need_tmp64 = 1;
6473 break;
6474
6475 case OP_IMUL:
6476 if (po->operand_cnt == 1 && po->operand[0].lmod == OPLM_DWORD)
6477 need_tmp64 = 1;
6478 break;
6479
6480 case OP_CALL:
6481 // note: resolved non-reg calls are OPF_DONE already
6482 pp = po->pp;
6483 ferr_assert(po, pp != NULL);
6484
6485 if (pp->is_unresolved) {
6486 int regmask_stack = 0;
6487
6488 if ((po->flags & OPF_TAIL) && g_func_pp->is_stdcall)
6489 pp_insert_stack_args(pp, g_func_pp->argc_stack);
6490 else {
6491 collect_call_args(po, i, opcnt, pp, &regmask, i + opcnt * 2);
6492
6493 // this is pretty rough guess:
6494 // see ecx and edx were pushed (and not their saved versions)
6495 for (arg = 0; arg < pp->argc; arg++) {
6496 if (pp->arg[arg].reg != NULL && !pp->arg[arg].is_saved)
6497 continue;
6498
6499 if (pp->arg[arg].push_ref_cnt == 0)
6500 ferr(po, "parsed_op missing for arg%d\n", arg);
6501 tmp_op = pp->arg[arg].push_refs[0];
6502 if (tmp_op->operand[0].type == OPT_REG)
6503 regmask_stack |= 1 << tmp_op->operand[0].reg;
6504 }
6505 }
6506
6507 // quick dumb check for potential reg-args
6508 for (j = i - 1; j >= 0 && ops[j].op == OP_MOV; j--)
6509 if (ops[j].operand[0].type == OPT_REG)
6510 regmask_stack &= ~(1 << ops[j].operand[0].reg);
6511
6512 if ((regmask_stack & (mxCX|mxDX)) != (mxCX|mxDX)
6513 && ((regmask | regmask_arg) & (mxCX|mxDX)))
6514 {
6515 if (pp->argc_stack != 0
6516 || ((regmask | regmask_arg) & (mxCX|mxDX)))
6517 {
6518 pp_insert_reg_arg(pp, "ecx");
6519 pp->is_fastcall = 1;
6520 regmask_init |= 1 << xCX;
6521 regmask |= 1 << xCX;
6522 }
6523 if (pp->argc_stack != 0
6524 || ((regmask | regmask_arg) & mxDX))
6525 {
6526 pp_insert_reg_arg(pp, "edx");
6527 regmask_init |= 1 << xDX;
6528 regmask |= 1 << xDX;
6529 }
6530 }
6531
6532 // note: __cdecl doesn't fall into is_unresolved category
6533 if (pp->argc_stack > 0)
6534 pp->is_stdcall = 1;
6535 }
6536 if (!(po->flags & OPF_TAIL)
6537 && !(g_sct_func_attr & SCTFA_NOWARN) && !g_nowarn_reguse)
6538 {
6539 // treat al write as overwrite to avoid many false positives
6540 if (IS(pp->ret_type.name, "void") || pp->ret_type.is_float) {
6541 find_next_read_reg(i + 1, opcnt, xAX, OPLM_BYTE,
6542 i + opcnt * 25, &j);
6543 if (j != -1) {
6544 fnote(po, "eax used after void/float ret call\n");
6545 fnote(&ops[j], "(used here)\n");
6546 }
6547 }
6548 if (!strstr(pp->ret_type.name, "int64")) {
6549 find_next_read_reg(i + 1, opcnt, xDX, OPLM_BYTE,
6550 i + opcnt * 26, &j);
6551 // indirect calls are often guessed, don't warn
6552 if (j != -1 && !IS_OP_INDIRECT_CALL(&ops[j])) {
6553 fnote(po, "edx used after 32bit ret call\n");
6554 fnote(&ops[j], "(used here)\n");
6555 }
6556 }
6557 j = 1;
6558 // msvc often relies on callee not modifying 'this'
6559 for (arg = 0; arg < pp->argc; arg++) {
6560 if (pp->arg[arg].reg && IS(pp->arg[arg].reg, "ecx")) {
6561 j = 0;
6562 break;
6563 }
6564 }
6565 if (j != 0) {
6566 find_next_read_reg(i + 1, opcnt, xCX, OPLM_BYTE,
6567 i + opcnt * 27, &j);
6568 if (j != -1 && !IS_OP_INDIRECT_CALL(&ops[j])) {
6569 fnote(po, "ecx used after call\n");
6570 fnote(&ops[j], "(used here)\n");
6571 }
6572 }
6573 }
6574 break;
6575
6576 case OP_MOV:
6577 if (po->operand[0].pp != NULL && po->operand[1].pp != NULL)
6578 {
6579 // <var> = offset <something>
6580 if ((po->operand[1].pp->is_func || po->operand[1].pp->is_fptr)
6581 && !IS_START(po->operand[1].name, "off_"))
6582 {
6583 if (!po->operand[0].pp->is_fptr)
6584 ferr(po, "%s not declared as fptr when it should be\n",
6585 po->operand[0].name);
6586 if (pp_cmp_func(po->operand[0].pp, po->operand[1].pp)) {
6587 pp_print(buf1, sizeof(buf1), po->operand[0].pp);
6588 pp_print(buf2, sizeof(buf2), po->operand[1].pp);
6589 fnote(po, "var: %s\n", buf1);
6590 fnote(po, "func: %s\n", buf2);
6591 ferr(po, "^ mismatch\n");
6592 }
6593 }
6594 }
6595 break;
6596
6597 case OP_DIV:
6598 case OP_IDIV:
6599 if (po->operand[0].lmod == OPLM_DWORD) {
6600 // 32bit division is common, look for it
6601 if (po->op == OP_DIV)
6602 ret = scan_for_reg_clear(i, xDX);
6603 else
6604 ret = scan_for_cdq_edx(i);
6605 if (ret >= 0)
6606 po->flags |= OPF_32BIT;
6607 else
6608 need_tmp64 = 1;
6609 }
6610 else
6611 need_tmp_var = 1;
6612 break;
6613
6614 case OP_CLD:
6615 po->flags |= OPF_RMD | OPF_DONE;
6616 break;
6617
6618 case OP_RCL:
6619 case OP_RCR:
6620 case OP_XCHG:
6621 need_tmp_var = 1;
6622 break;
6623
6624 case OP_FLD:
6625 if (po->operand[0].lmod == OPLM_QWORD)
6626 need_double = 1;
6627 break;
6628
6629 case OP_RDTSC:
6630 case OPP_ALLSHL:
6631 case OPP_ALLSHR:
6632 need_tmp64 = 1;
6633 break;
6634
6635 case OPP_FTOL:
6636 find_next_read_reg(i + 1, opcnt, xDX, OPLM_DWORD,
6637 i + opcnt * 18, &j);
6638 if (j == -1)
6639 po->flags |= OPF_32BIT;
6640 break;
6641
6642 default:
6643 break;
6644 }
6645 }
6646
6647 // pass8: sync all push arg numbers
6648 // some calls share args and not all of them
6649 // (there's only partial intersection)
6650 do {
6651 int changed, argnum, arggrp;
6652
6653 found = 0;
6654 for (i = 0; i < opcnt; i++)
6655 {
6656 po = &ops[i];
6657 if ((po->flags & (OPF_RMD|OPF_DONE)) || po->op != OP_CALL)
6658 continue;
6659
6660 pp = po->pp;
6661 arggrp = 0;
6662 do {
6663 changed = 0;
6664 for (arg = argnum = 0; arg < pp->argc; arg++) {
6665 if (pp->arg[arg].reg != NULL)
6666 continue;
6667 if (pp->arg[arg].is_saved)
6668 changed |= sync_argnum(pp, arg, &argnum, &arggrp);
6669 argnum++;
6670 }
6671 found |= changed;
6672 }
6673 while (changed);
6674
6675 if (argnum > 32)
6676 ferr(po, "too many args or looping in graph\n");
6677 }
6678 }
6679 while (found);
6680
6681 // pass9: final adjustments
6682 for (i = 0; i < opcnt; i++)
6683 {
6684 po = &ops[i];
6685 if (po->flags & (OPF_RMD|OPF_DONE))
6686 continue;
6687
6688 if (po->op != OP_FST && po->p_argnum > 0)
6689 save_arg_vars[po->p_arggrp] |= 1 << (po->p_argnum - 1);
6690
6691 // correct for "full stack" mode late enable
6692 if ((po->flags & (OPF_PPUSH|OPF_FPOP|OPF_FPOPP))
6693 && need_float_stack)
6694 po->flags |= OPF_FSHIFT;
6695 }
6696
6697 float_type = need_double ? "double" : "float";
6698 float_st0 = need_float_stack ? "f_st[f_stp & 7]" : "f_st0";
6699 float_st1 = need_float_stack ? "f_st[(f_stp + 1) & 7]" : "f_st1";
6700
6701 // output starts here
6702
6703 if (g_seh_found)
6704 fprintf(fout, "// had SEH\n");
6705
6706 // define userstack size
6707 if (g_func_pp->is_userstack) {
6708 fprintf(fout, "#ifndef US_SZ_%s\n", g_func_pp->name);
6709 fprintf(fout, "#define US_SZ_%s USERSTACK_SIZE\n", g_func_pp->name);
6710 fprintf(fout, "#endif\n");
6711 }
6712
6713 // the function itself
6714 ferr_assert(ops, !g_func_pp->is_fptr);
6715 output_pp(fout, g_func_pp,
6716 (g_ida_func_attr & IDAFA_NORETURN) ? OPP_FORCE_NORETURN : 0);
6717 fprintf(fout, "\n{\n");
6718
6719 // declare indirect functions
6720 for (i = 0; i < opcnt; i++) {
6721 po = &ops[i];
6722 if (po->flags & OPF_RMD)
6723 continue;
6724
6725 if (po->op == OP_CALL) {
6726 pp = po->pp;
6727 if (pp == NULL)
6728 ferr(po, "NULL pp\n");
6729
6730 if (pp->is_fptr && !(pp->name[0] != 0 && pp->is_arg)) {
6731 if (pp->name[0] != 0) {
6732 if (IS_START(pp->name, "guess"))
6733 pp->is_guessed = 1;
6734
6735 memmove(pp->name + 2, pp->name, strlen(pp->name) + 1);
6736 memcpy(pp->name, "i_", 2);
6737
6738 // might be declared already
6739 found = 0;
6740 for (j = 0; j < i; j++) {
6741 if (ops[j].op == OP_CALL && (pp_tmp = ops[j].pp)) {
6742 if (pp_tmp->is_fptr && IS(pp->name, pp_tmp->name)) {
6743 found = 1;
6744 break;
6745 }
6746 }
6747 }
6748 if (found)
6749 continue;
6750 }
6751 else
6752 snprintf(pp->name, sizeof(pp->name), "icall%d", i);
6753
6754 fprintf(fout, " ");
6755 output_pp(fout, pp, OPP_SIMPLE_ARGS);
6756 fprintf(fout, ";\n");
6757 }
6758 }
6759 }
6760
6761 // output LUTs/jumptables
6762 for (i = 0; i < g_func_pd_cnt; i++) {
6763 pd = &g_func_pd[i];
6764 fprintf(fout, " static const ");
6765 if (pd->type == OPT_OFFSET) {
6766 fprintf(fout, "void *jt_%s[] =\n { ", pd->label);
6767
6768 for (j = 0; j < pd->count; j++) {
6769 if (j > 0)
6770 fprintf(fout, ", ");
6771 fprintf(fout, "&&%s", pd->d[j].u.label);
6772 }
6773 }
6774 else {
6775 fprintf(fout, "%s %s[] =\n { ",
6776 lmod_type_u(ops, pd->lmod), pd->label);
6777
6778 for (j = 0; j < pd->count; j++) {
6779 if (j > 0)
6780 fprintf(fout, ", ");
6781 fprintf(fout, "%u", pd->d[j].u.val);
6782 }
6783 }
6784 fprintf(fout, " };\n");
6785 had_decl = 1;
6786 }
6787
6788 // declare stack frame, va_arg
6789 if (g_stack_fsz) {
6790 if (stack_fsz_adj)
6791 fprintf(fout, " // stack_fsz_adj %d\n", stack_fsz_adj);
6792
6793 fprintf(fout, " union { u32 d[%d];", (g_stack_fsz + 3) / 4);
6794 if (g_func_lmods & (1 << OPLM_WORD))
6795 fprintf(fout, " u16 w[%d];", (g_stack_fsz + 1) / 2);
6796 if (g_func_lmods & (1 << OPLM_BYTE))
6797 fprintf(fout, " u8 b[%d];", g_stack_fsz);
6798 if (g_func_lmods & (1 << OPLM_QWORD))
6799 fprintf(fout, " double q[%d];", (g_stack_fsz + 7) / 8);
6800
6801 if (stack_align > 8)
6802 ferr(ops, "unhandled stack align of %d\n", stack_align);
6803 else if (stack_align == 8)
6804 fprintf(fout, " u64 align;");
6805 fprintf(fout, " } sf;\n");
6806 had_decl = 1;
6807 }
6808
6809 if ((g_sct_func_attr & SCTFA_ARGFRAME) && g_func_pp->argc_stack) {
6810 fprintf(fout, " struct { u32 ");
6811 for (i = j = 0; i < g_func_pp->argc; i++) {
6812 if (g_func_pp->arg[i].reg != NULL)
6813 continue;
6814 if (j++ != 0)
6815 fprintf(fout, ", ");
6816 fprintf(fout, "a%d", i + 1);
6817 }
6818 fprintf(fout, "; } af = {\n ");
6819 for (i = j = 0; i < g_func_pp->argc; i++) {
6820 if (g_func_pp->arg[i].reg != NULL)
6821 continue;
6822 if (j++ != 0)
6823 fprintf(fout, ", ");
6824 if (g_func_pp->arg[i].type.is_ptr)
6825 fprintf(fout, "(u32)");
6826 fprintf(fout, "a%d", i + 1);
6827 }
6828 fprintf(fout, "\n };\n");
6829 }
6830
6831 if (g_func_pp->is_userstack) {
6832 fprintf(fout, " u32 fake_sf[US_SZ_%s / 4];\n", g_func_pp->name);
6833 fprintf(fout, " u32 *esp = &fake_sf[sizeof(fake_sf) / 4];\n");
6834 had_decl = 1;
6835 }
6836
6837 if (g_func_pp->is_vararg) {
6838 fprintf(fout, " va_list ap;\n");
6839 had_decl = 1;
6840 }
6841
6842 // declare arg-registers
6843 for (i = 0; i < g_func_pp->argc; i++) {
6844 if (g_func_pp->arg[i].reg != NULL) {
6845 reg = char_array_i(regs_r32,
6846 ARRAY_SIZE(regs_r32), g_func_pp->arg[i].reg);
6847 if (regmask & (1 << reg)) {
6848 if (g_func_pp->arg[i].type.is_retreg)
6849 fprintf(fout, " u32 %s = *r_%s;\n",
6850 g_func_pp->arg[i].reg, g_func_pp->arg[i].reg);
6851 else
6852 fprintf(fout, " u32 %s = (u32)a%d;\n",
6853 g_func_pp->arg[i].reg, i + 1);
6854 }
6855 else {
6856 if (g_func_pp->arg[i].type.is_retreg)
6857 ferr(ops, "retreg '%s' is unused?\n",
6858 g_func_pp->arg[i].reg);
6859 fprintf(fout, " // %s = a%d; // unused\n",
6860 g_func_pp->arg[i].reg, i + 1);
6861 }
6862 had_decl = 1;
6863 }
6864 }
6865
6866 // declare normal registers
6867 regmask_now = regmask & ~regmask_arg & ~g_regmask_rm;
6868 regmask_now &= ~(1 << xSP);
6869 if (regmask_now & 0x00ff) {
6870 for (reg = 0; reg < 8; reg++) {
6871 if (regmask_now & (1 << reg)) {
6872 fprintf(fout, " u32 %s", regs_r32[reg]);
6873 if (regmask_init & (1 << reg))
6874 fprintf(fout, " = 0");
6875 fprintf(fout, ";\n");
6876 had_decl = 1;
6877 }
6878 }
6879 }
6880 // ... mmx
6881 if (regmask_now & 0xff00) {
6882 for (reg = 8; reg < 16; reg++) {
6883 if (regmask_now & (1 << reg)) {
6884 fprintf(fout, " mmxr %s", regs_r32[reg]);
6885 if (regmask_init & (1 << reg))
6886 fprintf(fout, " = { 0, }");
6887 fprintf(fout, ";\n");
6888 had_decl = 1;
6889 }
6890 }
6891 }
6892 // ... x87
6893 if (need_float_stack) {
6894 fprintf(fout, " %s f_st[8];\n", float_type);
6895 fprintf(fout, " int f_stp = 0;\n");
6896 had_decl = 1;
6897 }
6898 else {
6899 if (regmask_now & 0xff0000) {
6900 for (reg = 16; reg < 24; reg++) {
6901 if (regmask_now & (1 << reg)) {
6902 fprintf(fout, " %s f_st%d", float_type, reg - 16);
6903 if (regmask_init & (1 << reg))
6904 fprintf(fout, " = 0");
6905 fprintf(fout, ";\n");
6906 had_decl = 1;
6907 }
6908 }
6909 }
6910 }
6911
6912 if (need_float_sw) {
6913 fprintf(fout, " u16 f_sw;\n");
6914 had_decl = 1;
6915 }
6916
6917 if (regmask_save) {
6918 for (reg = 0; reg < 8; reg++) {
6919 if (regmask_save & (1 << reg)) {
6920 fprintf(fout, " u32 s_%s;\n", regs_r32[reg]);
6921 had_decl = 1;
6922 }
6923 }
6924 }
6925
6926 for (i = 0; i < ARRAY_SIZE(save_arg_vars); i++) {
6927 if (save_arg_vars[i] == 0)
6928 continue;
6929 for (reg = 0; reg < 32; reg++) {
6930 if (save_arg_vars[i] & (1 << reg)) {
6931 fprintf(fout, " u32 %s;\n",
6932 saved_arg_name(buf1, sizeof(buf1), i, reg + 1));
6933 had_decl = 1;
6934 }
6935 }
6936 }
6937
6938 if (regmask_ffca) {
6939 for (reg = 0; reg < 32; reg++) {
6940 if (regmask_ffca & (1 << reg)) {
6941 fprintf(fout, " %s fs_%d;\n", float_type, reg + 1);
6942 had_decl = 1;
6943 }
6944 }
6945 }
6946
6947 // declare push-pop temporaries
6948 if (regmask_pp) {
6949 for (reg = 0; reg < 8; reg++) {
6950 if (regmask_pp & (1 << reg)) {
6951 fprintf(fout, " u32 pp_%s;\n", regs_r32[reg]);
6952 had_decl = 1;
6953 }
6954 }
6955 }
6956
6957 if (cond_vars) {
6958 for (i = 0; i < 8; i++) {
6959 if (cond_vars & (1 << i)) {
6960 fprintf(fout, " u32 cond_%s;\n", parsed_flag_op_names[i]);
6961 had_decl = 1;
6962 }
6963 }
6964 }
6965
6966 if (need_tmp_var) {
6967 fprintf(fout, " u32 tmp;\n");
6968 had_decl = 1;
6969 }
6970
6971 if (need_tmp64) {
6972 fprintf(fout, " u64 tmp64;\n");
6973 had_decl = 1;
6974 }
6975
6976 if (had_decl)
6977 fprintf(fout, "\n");
6978
6979 // do stack clear, if needed
6980 if (g_sct_func_attr & SCTFA_CLEAR_SF) {
6981 fprintf(fout, " ");
6982 if (g_stack_clear_len != 0) {
6983 if (g_stack_clear_len <= 4) {
6984 for (i = 0; i < g_stack_clear_len; i++)
6985 fprintf(fout, "sf.d[%d] = ", g_stack_clear_start + i);
6986 fprintf(fout, "0;\n");
6987 }
6988 else {
6989 fprintf(fout, "memset(&sf[%d], 0, %d);\n",
6990 g_stack_clear_start, g_stack_clear_len * 4);
6991 }
6992 }
6993 else
6994 fprintf(fout, "memset(&sf, 0, sizeof(sf));\n");
6995 }
6996
6997 if (g_func_pp->is_vararg) {
6998 if (g_func_pp->argc_stack == 0)
6999 ferr(ops, "vararg func without stack args?\n");
7000 fprintf(fout, " va_start(ap, a%d);\n", g_func_pp->argc);
7001 }
7002
7003 // output ops
7004 for (i = 0; i < opcnt; i++)
7005 {
7006 if (g_labels[i] != NULL) {
7007 fprintf(fout, "\n%s:\n", g_labels[i]);
7008 label_pending = 1;
7009
7010 delayed_flag_op = NULL;
7011 last_arith_dst = NULL;
7012 }
7013
7014 po = &ops[i];
7015 if (po->flags & OPF_RMD)
7016 continue;
7017
7018 lock_handled = 0;
7019 no_output = 0;
7020
7021 #define assert_operand_cnt(n_) \
7022 if (po->operand_cnt != n_) \
7023 ferr(po, "operand_cnt is %d/%d\n", po->operand_cnt, n_)
7024
7025 // conditional/flag using op?
7026 if (po->flags & OPF_CC)
7027 {
7028 int is_delayed = 0;
7029
7030 tmp_op = po->datap;
7031
7032 // we go through all this trouble to avoid using parsed_flag_op,
7033 // which makes generated code much nicer
7034 if (delayed_flag_op != NULL)
7035 {
7036 out_cmp_test(buf1, sizeof(buf1), delayed_flag_op,
7037 po->pfo, po->pfo_inv);
7038 is_delayed = 1;
7039 }
7040 else if (last_arith_dst != NULL
7041 && (po->pfo == PFO_Z || po->pfo == PFO_S || po->pfo == PFO_P
7042 || (tmp_op && (tmp_op->op == OP_AND || tmp_op->op == OP_OR))
7043 ))
7044 {
7045 struct parsed_op *po_arith = (void *)((char *)last_arith_dst
7046 - offsetof(struct parsed_op, operand[0]));
7047 ferr_assert(po, &ops[po_arith - ops] == po_arith);
7048 out_src_opr_u32(buf3, sizeof(buf3), po_arith, last_arith_dst);
7049 out_test_for_cc(buf1, sizeof(buf1), po, po->pfo, po->pfo_inv,
7050 last_arith_dst->lmod, buf3);
7051 is_delayed = 1;
7052 }
7053 else if (tmp_op != NULL) {
7054 // use preprocessed flag calc results
7055 if (!(tmp_op->pfomask & (1 << po->pfo)))
7056 ferr(po, "not prepared for pfo %d\n", po->pfo);
7057
7058 // note: pfo_inv was not yet applied
7059 snprintf(buf1, sizeof(buf1), "(%scond_%s)",
7060 po->pfo_inv ? "!" : "", parsed_flag_op_names[po->pfo]);
7061 }
7062 else {
7063 ferr(po, "all methods of finding comparison failed\n");
7064 }
7065
7066 if (po->flags & OPF_JMP) {
7067 fprintf(fout, " if %s", buf1);
7068 }
7069 else if (po->op == OP_RCL || po->op == OP_RCR
7070 || po->op == OP_ADC || po->op == OP_SBB)
7071 {
7072 if (is_delayed)
7073 fprintf(fout, " cond_%s = %s;\n",
7074 parsed_flag_op_names[po->pfo], buf1);
7075 }
7076 else if (po->flags & OPF_DATA) { // SETcc
7077 out_dst_opr(buf2, sizeof(buf2), po, &po->operand[0]);
7078 fprintf(fout, " %s = %s;", buf2, buf1);
7079 }
7080 else {
7081 ferr(po, "unhandled conditional op\n");
7082 }
7083 }
7084
7085 pfomask = po->pfomask;
7086
7087 switch (po->op)
7088 {
7089 case OP_MOV:
7090 assert_operand_cnt(2);
7091 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7092 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7093 default_cast_to(buf3, sizeof(buf3), &po->operand[0]);
7094 fprintf(fout, " %s = %s;", buf1,
7095 out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
7096 buf3, 0));
7097 break;
7098
7099 case OP_LEA:
7100 assert_operand_cnt(2);
7101 po->operand[1].lmod = OPLM_DWORD; // always
7102 fprintf(fout, " %s = %s;",
7103 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7104 out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
7105 NULL, 1));
7106 break;
7107
7108 case OP_MOVZX:
7109 assert_operand_cnt(2);
7110 fprintf(fout, " %s = %s;",
7111 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7112 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
7113 break;
7114
7115 case OP_MOVSX:
7116 assert_operand_cnt(2);
7117 switch (po->operand[1].lmod) {
7118 case OPLM_BYTE:
7119 strcpy(buf3, "(s8)");
7120 break;
7121 case OPLM_WORD:
7122 strcpy(buf3, "(s16)");
7123 break;
7124 default:
7125 ferr(po, "invalid src lmod: %d\n", po->operand[1].lmod);
7126 }
7127 fprintf(fout, " %s = %s;",
7128 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7129 out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
7130 buf3, 0));
7131 break;
7132
7133 case OP_XCHG:
7134 assert_operand_cnt(2);
7135 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7136 fprintf(fout, " tmp = %s;",
7137 out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], "", 0));
7138 fprintf(fout, " %s = %s;",
7139 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7140 out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
7141 default_cast_to(buf3, sizeof(buf3), &po->operand[0]), 0));
7142 fprintf(fout, " %s = %stmp;",
7143 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[1]),
7144 default_cast_to(buf3, sizeof(buf3), &po->operand[1]));
7145 snprintf(g_comment, sizeof(g_comment), "xchg");
7146 break;
7147
7148 case OP_NOT:
7149 assert_operand_cnt(1);
7150 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7151 fprintf(fout, " %s = ~%s;", buf1, buf1);
7152 break;
7153
7154 case OP_XLAT:
7155 assert_operand_cnt(2);
7156 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7157 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
7158 fprintf(fout, " %s = *(u8 *)(%s + %s);", buf1, buf2, buf1);
7159 strcpy(g_comment, "xlat");
7160 break;
7161
7162 case OP_CDQ:
7163 assert_operand_cnt(2);
7164 fprintf(fout, " %s = (s32)%s >> 31;",
7165 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7166 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
7167 strcpy(g_comment, "cdq");
7168 break;
7169
7170 case OP_BSWAP:
7171 assert_operand_cnt(1);
7172 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7173 fprintf(fout, " %s = __builtin_bswap32(%s);", buf1, buf1);
7174 break;
7175
7176 case OP_LODS:
7177 if (po->flags & OPF_REP) {
7178 assert_operand_cnt(3);
7179 // hmh..
7180 ferr(po, "TODO\n");
7181 }
7182 else {
7183 assert_operand_cnt(2);
7184 fprintf(fout, " %s = %sesi; esi %c= %d;",
7185 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[1]),
7186 lmod_cast_u_ptr(po, po->operand[1].lmod),
7187 (po->flags & OPF_DF) ? '-' : '+',
7188 lmod_bytes(po, po->operand[1].lmod));
7189 strcpy(g_comment, "lods");
7190 }
7191 break;
7192
7193 case OP_STOS:
7194 if (po->flags & OPF_REP) {
7195 assert_operand_cnt(3);
7196 fprintf(fout, " for (; ecx != 0; ecx--, edi %c= %d)\n",
7197 (po->flags & OPF_DF) ? '-' : '+',
7198 lmod_bytes(po, po->operand[1].lmod));
7199 fprintf(fout, " %sedi = eax;\n",
7200 lmod_cast_u_ptr(po, po->operand[1].lmod));
7201 fprintf(fout, " barrier();");
7202 strcpy(g_comment, "^ rep stos");
7203 }
7204 else {
7205 assert_operand_cnt(2);
7206 fprintf(fout, " %sedi = eax; edi %c= %d;",
7207 lmod_cast_u_ptr(po, po->operand[1].lmod),
7208 (po->flags & OPF_DF) ? '-' : '+',
7209 lmod_bytes(po, po->operand[1].lmod));
7210 strcpy(g_comment, "stos");
7211 }
7212 break;
7213
7214 case OP_MOVS:
7215 j = lmod_bytes(po, po->operand[0].lmod);
7216 strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod));
7217 l = (po->flags & OPF_DF) ? '-' : '+';
7218 if (po->flags & OPF_REP) {
7219 assert_operand_cnt(3);
7220 fprintf(fout,
7221 " for (; ecx != 0; ecx--, edi %c= %d, esi %c= %d)\n",
7222 l, j, l, j);
7223 fprintf(fout,
7224 " %sedi = %sesi;\n", buf1, buf1);
7225 // this can overwrite many variables
7226 fprintf(fout, " barrier();");
7227 strcpy(g_comment, "^ rep movs");
7228 }
7229 else {
7230 assert_operand_cnt(2);
7231 fprintf(fout, " %sedi = %sesi; edi %c= %d; esi %c= %d;",
7232 buf1, buf1, l, j, l, j);
7233 strcpy(g_comment, "movs");
7234 }
7235 break;
7236
7237 case OP_CMPS:
7238 // repe ~ repeat while ZF=1
7239 j = lmod_bytes(po, po->operand[0].lmod);
7240 strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod));
7241 l = (po->flags & OPF_DF) ? '-' : '+';
7242 if (po->flags & OPF_REP) {
7243 assert_operand_cnt(3);
7244 fprintf(fout,
7245 " while (ecx != 0) {\n");
7246 if (pfomask & (1 << PFO_C)) {
7247 // ugh..
7248 fprintf(fout,
7249 " cond_c = %sesi < %sedi;\n", buf1, buf1);
7250 pfomask &= ~(1 << PFO_C);
7251 }
7252 fprintf(fout,
7253 " cond_z = (%sesi == %sedi); esi %c= %d, edi %c= %d;\n",
7254 buf1, buf1, l, j, l, j);
7255 fprintf(fout,
7256 " ecx--;\n"
7257 " if (cond_z %s 0) break;\n",
7258 (po->flags & OPF_REPZ) ? "==" : "!=");
7259 fprintf(fout,
7260 " }");
7261 snprintf(g_comment, sizeof(g_comment), "rep%s cmps",
7262 (po->flags & OPF_REPZ) ? "e" : "ne");
7263 }
7264 else {
7265 assert_operand_cnt(2);
7266 fprintf(fout,
7267 " cond_z = (%sesi == %sedi); esi %c= %d; edi %c= %d;",
7268 buf1, buf1, l, j, l, j);
7269 strcpy(g_comment, "cmps");
7270 }
7271 pfomask &= ~(1 << PFO_Z);
7272 last_arith_dst = NULL;
7273 delayed_flag_op = NULL;
7274 break;
7275
7276 case OP_SCAS:
7277 // only does ZF (for now)
7278 // repe ~ repeat while ZF=1
7279 j = lmod_bytes(po, po->operand[1].lmod);
7280 l = (po->flags & OPF_DF) ? '-' : '+';
7281 if (po->flags & OPF_REP) {
7282 assert_operand_cnt(3);
7283 fprintf(fout,
7284 " while (ecx != 0) {\n");
7285 fprintf(fout,
7286 " cond_z = (%seax == %sedi); edi %c= %d;\n",
7287 lmod_cast_u(po, po->operand[1].lmod),
7288 lmod_cast_u_ptr(po, po->operand[1].lmod), l, j);
7289 fprintf(fout,
7290 " ecx--;\n"
7291 " if (cond_z %s 0) break;\n",
7292 (po->flags & OPF_REPZ) ? "==" : "!=");
7293 fprintf(fout,
7294 " }");
7295 snprintf(g_comment, sizeof(g_comment), "rep%s scas",
7296 (po->flags & OPF_REPZ) ? "e" : "ne");
7297 }
7298 else {
7299 assert_operand_cnt(2);
7300 fprintf(fout, " cond_z = (%seax == %sedi); edi %c= %d;",
7301 lmod_cast_u(po, po->operand[1].lmod),
7302 lmod_cast_u_ptr(po, po->operand[1].lmod), l, j);
7303 strcpy(g_comment, "scas");
7304 }
7305 pfomask &= ~(1 << PFO_Z);
7306 last_arith_dst = NULL;
7307 delayed_flag_op = NULL;
7308 break;
7309
7310 case OP_RDTSC:
7311 fprintf(fout, " tmp64 = ext_rdtsc();\n");
7312 fprintf(fout, " edx = tmp64 >> 32;\n");
7313 fprintf(fout, " eax = tmp64;");
7314 break;
7315
7316 case OP_CPUID:
7317 fprintf(fout, " ext_cpuid(&eax, &ebx, &ecx, &edx);");
7318 break;
7319
7320 // arithmetic w/flags
7321 case OP_AND:
7322 if (po->operand[1].type == OPT_CONST && !po->operand[1].val)
7323 goto dualop_arith_const;
7324 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7325 goto dualop_arith;
7326
7327 case OP_OR:
7328 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7329 if (po->operand[1].type == OPT_CONST) {
7330 j = lmod_bytes(po, po->operand[0].lmod);
7331 if (((1ull << j * 8) - 1) == po->operand[1].val)
7332 goto dualop_arith_const;
7333 }
7334 goto dualop_arith;
7335
7336 dualop_arith:
7337 assert_operand_cnt(2);
7338 fprintf(fout, " %s %s= %s;",
7339 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7340 op_to_c(po),
7341 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
7342 output_std_flags(fout, po, &pfomask, buf1);
7343 last_arith_dst = &po->operand[0];
7344 delayed_flag_op = NULL;
7345 break;
7346
7347 dualop_arith_const:
7348 // and 0, or ~0 used instead mov
7349 assert_operand_cnt(2);
7350 fprintf(fout, " %s = %s;",
7351 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7352 out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
7353 default_cast_to(buf3, sizeof(buf3), &po->operand[0]), 0));
7354 output_std_flags(fout, po, &pfomask, buf1);
7355 last_arith_dst = &po->operand[0];
7356 delayed_flag_op = NULL;
7357 break;
7358
7359 case OP_SHL:
7360 case OP_SHR:
7361 assert_operand_cnt(2);
7362 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7363 if (pfomask & (1 << PFO_C)) {
7364 if (po->operand[1].type == OPT_CONST) {
7365 l = lmod_bytes(po, po->operand[0].lmod) * 8;
7366 j = po->operand[1].val;
7367 j %= l;
7368 if (j != 0) {
7369 if (po->op == OP_SHL)
7370 j = l - j;
7371 else
7372 j -= 1;
7373 fprintf(fout, " cond_c = (%s >> %d) & 1;\n",
7374 buf1, j);
7375 }
7376 else
7377 ferr(po, "zero shift?\n");
7378 }
7379 else
7380 ferr(po, "TODO\n");
7381 pfomask &= ~(1 << PFO_C);
7382 }
7383 fprintf(fout, " %s %s= %s", buf1, op_to_c(po),
7384 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
7385 if (po->operand[1].type != OPT_CONST)
7386 fprintf(fout, " & 0x1f");
7387 fprintf(fout, ";");
7388 output_std_flags(fout, po, &pfomask, buf1);
7389 last_arith_dst = &po->operand[0];
7390 delayed_flag_op = NULL;
7391 break;
7392
7393 case OP_SAR:
7394 assert_operand_cnt(2);
7395 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7396 fprintf(fout, " %s = %s%s >> %s;", buf1,
7397 lmod_cast_s(po, po->operand[0].lmod), buf1,
7398 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
7399 output_std_flags(fout, po, &pfomask, buf1);
7400 last_arith_dst = &po->operand[0];
7401 delayed_flag_op = NULL;
7402 break;
7403
7404 case OP_SHLD:
7405 case OP_SHRD:
7406 assert_operand_cnt(3);
7407 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7408 l = lmod_bytes(po, po->operand[0].lmod) * 8;
7409 out_src_opr_u32(buf3, sizeof(buf3), po, &po->operand[2]);
7410 if (po->operand[2].type != OPT_CONST) {
7411 // no handling for "undefined" case, hopefully not needed
7412 snprintf(buf2, sizeof(buf2), "(%s & 0x1f)", buf3);
7413 strcpy(buf3, buf2);
7414 }
7415 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
7416 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7417 if (po->op == OP_SHLD) {
7418 fprintf(fout, " %s <<= %s; %s |= %s >> (%d - %s);",
7419 buf1, buf3, buf1, buf2, l, buf3);
7420 strcpy(g_comment, "shld");
7421 }
7422 else {
7423 fprintf(fout, " %s >>= %s; %s |= %s << (%d - %s);",
7424 buf1, buf3, buf1, buf2, l, buf3);
7425 strcpy(g_comment, "shrd");
7426 }
7427 output_std_flags(fout, po, &pfomask, buf1);
7428 last_arith_dst = &po->operand[0];
7429 delayed_flag_op = NULL;
7430 break;
7431
7432 case OP_ROL:
7433 case OP_ROR:
7434 assert_operand_cnt(2);
7435 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7436 if (po->operand[1].type == OPT_CONST) {
7437 j = po->operand[1].val;
7438 j %= lmod_bytes(po, po->operand[0].lmod) * 8;
7439 fprintf(fout, po->op == OP_ROL ?
7440 " %s = (%s << %d) | (%s >> %d);" :
7441 " %s = (%s >> %d) | (%s << %d);",
7442 buf1, buf1, j, buf1,
7443 lmod_bytes(po, po->operand[0].lmod) * 8 - j);
7444 }
7445 else
7446 ferr(po, "TODO\n");
7447 output_std_flags(fout, po, &pfomask, buf1);
7448 last_arith_dst = &po->operand[0];
7449 delayed_flag_op = NULL;
7450 break;
7451
7452 case OP_RCL:
7453 case OP_RCR:
7454 assert_operand_cnt(2);
7455 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7456 l = lmod_bytes(po, po->operand[0].lmod) * 8;
7457 if (po->operand[1].type == OPT_CONST) {
7458 j = po->operand[1].val % l;
7459 if (j == 0)
7460 ferr(po, "zero rotate\n");
7461 fprintf(fout, " tmp = (%s >> %d) & 1;\n",
7462 buf1, (po->op == OP_RCL) ? (l - j) : (j - 1));
7463 if (po->op == OP_RCL) {
7464 fprintf(fout,
7465 " %s = (%s << %d) | (cond_c << %d)",
7466 buf1, buf1, j, j - 1);
7467 if (j != 1)
7468 fprintf(fout, " | (%s >> %d)", buf1, l + 1 - j);
7469 }
7470 else {
7471 fprintf(fout,
7472 " %s = (%s >> %d) | (cond_c << %d)",
7473 buf1, buf1, j, l - j);
7474 if (j != 1)
7475 fprintf(fout, " | (%s << %d)", buf1, l + 1 - j);
7476 }
7477 fprintf(fout, ";\n");
7478 fprintf(fout, " cond_c = tmp;");
7479 }
7480 else
7481 ferr(po, "TODO\n");
7482 strcpy(g_comment, (po->op == OP_RCL) ? "rcl" : "rcr");
7483 output_std_flags(fout, po, &pfomask, buf1);
7484 last_arith_dst = &po->operand[0];
7485 delayed_flag_op = NULL;
7486 break;
7487
7488 case OP_XOR:
7489 assert_operand_cnt(2);
7490 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7491 if (IS(opr_name(po, 0), opr_name(po, 1))) {
7492 // special case for XOR
7493 int z = PFOB_O | PFOB_C | PFOB_S | (1 << PFO_L);
7494 for (j = 0; j <= PFO_LE; j++) {
7495 if (pfomask & (1 << j)) {
7496 fprintf(fout, " cond_%s = %d;\n",
7497 parsed_flag_op_names[j], (1 << j) & z ? 0 : 1);
7498 pfomask &= ~(1 << j);
7499 }
7500 }
7501 fprintf(fout, " %s = 0;",
7502 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]));
7503 last_arith_dst = &po->operand[0];
7504 delayed_flag_op = NULL;
7505 break;
7506 }
7507 goto dualop_arith;
7508
7509 case OP_ADD:
7510 assert_operand_cnt(2);
7511 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7512 if (pfomask & (1 << PFO_C)) {
7513 out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
7514 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
7515 if (po->operand[0].lmod == OPLM_DWORD) {
7516 fprintf(fout, " tmp64 = (u64)%s + %s;\n", buf1, buf2);
7517 fprintf(fout, " cond_c = tmp64 >> 32;\n");
7518 fprintf(fout, " %s = (u32)tmp64;",
7519 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]));
7520 strcat(g_comment, " add64");
7521 }
7522 else {
7523 fprintf(fout, " cond_c = ((u32)%s + %s) >> %d;\n",
7524 buf1, buf2, lmod_bytes(po, po->operand[0].lmod) * 8);
7525 fprintf(fout, " %s += %s;",
7526 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
7527 buf2);
7528 }
7529 pfomask &= ~(1 << PFO_C);
7530 output_std_flags(fout, po, &pfomask, buf1);
7531 last_arith_dst = &po->operand[0];
7532 delayed_flag_op = NULL;
7533 break;
7534 }
7535 if (pfomask & (1 << PFO_LE)) {
7536 out_cmp_for_cc(buf1, sizeof(buf1), po, PFO_LE, 0, 1);
7537 fprintf(fout, " cond_%s = %s;\n",
7538 parsed_flag_op_names[PFO_LE], buf1);
7539 pfomask &= ~(1 << PFO_LE);
7540 }
7541 goto dualop_arith;
7542
7543 case OP_SUB:
7544 assert_operand_cnt(2);
7545 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7546 if (pfomask & ~((1 << PFO_Z) | (1 << PFO_S))) {
7547 for (j = 0; j <= PFO_LE; j++) {
7548 if (!(pfomask & (1 << j)))
7549 continue;
7550 if (j == PFO_Z || j == PFO_S)
7551 continue;
7552
7553 out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0, 0);
7554 fprintf(fout, " cond_%s = %s;\n",
7555 parsed_flag_op_names[j], buf1);
7556 pfomask &= ~(1 << j);
7557 }
7558 }
7559 goto dualop_arith;
7560
7561 case OP_ADC:
7562 case OP_SBB:
7563 assert_operand_cnt(2);
7564 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7565 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7566 if (po->op == OP_SBB
7567 && IS(po->operand[0].name, po->operand[1].name))
7568 {
7569 // avoid use of unitialized var
7570 fprintf(fout, " %s = -cond_c;", buf1);
7571 // carry remains what it was
7572 pfomask &= ~(1 << PFO_C);
7573 }
7574 else {
7575 fprintf(fout, " %s %s= %s + cond_c;", buf1, op_to_c(po),
7576 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
7577 }
7578 output_std_flags(fout, po, &pfomask, buf1);
7579 last_arith_dst = &po->operand[0];
7580 delayed_flag_op = NULL;
7581 break;
7582
7583 case OP_BSF:
7584 case OP_BSR:
7585 // on SKL, if src is 0, dst is left unchanged
7586 assert_operand_cnt(2);
7587 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7588 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
7589 output_std_flag_z(fout, po, &pfomask, buf2);
7590 if (po->op == OP_BSF)
7591 snprintf(buf3, sizeof(buf3), "__builtin_ffs(%s) - 1", buf2);
7592 else
7593 snprintf(buf3, sizeof(buf3), "31 - __builtin_clz(%s)", buf2);
7594 fprintf(fout, " if (%s) %s = %s;", buf2, buf1, buf3);
7595 last_arith_dst = &po->operand[0];
7596 delayed_flag_op = NULL;
7597 strcat(g_comment, po->op == OP_BSF ? " bsf" : " bsr");
7598 break;
7599
7600 case OP_DEC:
7601 if (pfomask & ~(PFOB_S | PFOB_S | PFOB_C)) {
7602 for (j = 0; j <= PFO_LE; j++) {
7603 if (!(pfomask & (1 << j)))
7604 continue;
7605 if (j == PFO_Z || j == PFO_S || j == PFO_C)
7606 continue;
7607
7608 out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0, 0);
7609 fprintf(fout, " cond_%s = %s;\n",
7610 parsed_flag_op_names[j], buf1);
7611 pfomask &= ~(1 << j);
7612 }
7613 }
7614 // fallthrough
7615
7616 case OP_INC:
7617 if (pfomask & (1 << PFO_C))
7618 // carry is unaffected by inc/dec.. wtf?
7619 ferr(po, "carry propagation needed\n");
7620
7621 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7622 if (po->operand[0].type == OPT_REG) {
7623 ferr_assert(po, !(po->flags & OPF_LOCK));
7624 strcpy(buf2, po->op == OP_INC ? "++" : "--");
7625 fprintf(fout, " %s%s;", buf1, buf2);
7626 }
7627 else if (po->flags & OPF_LOCK) {
7628 out_src_opr(buf2, sizeof(buf2), po, &po->operand[0], "", 1);
7629 fprintf(fout, " __sync_fetch_and_%s((%s *)(%s), 1);",
7630 po->op == OP_INC ? "add" : "sub",
7631 lmod_type_u(po, po->operand[0].lmod), buf2);
7632 strcat(g_comment, " lock");
7633 lock_handled = 1;
7634 }
7635 else {
7636 strcpy(buf2, po->op == OP_INC ? "+" : "-");
7637 fprintf(fout, " %s %s= 1;", buf1, buf2);
7638 }
7639 output_std_flags(fout, po, &pfomask, buf1);
7640 last_arith_dst = &po->operand[0];
7641 delayed_flag_op = NULL;
7642 break;
7643
7644 case OP_NEG:
7645 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
7646 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[0]);
7647 fprintf(fout, " %s = -%s%s;", buf1,
7648 lmod_cast_s(po, po->operand[0].lmod), buf2);
7649 last_arith_dst = &po->operand[0];
7650 delayed_flag_op = NULL;
7651 if (pfomask & PFOB_C) {
7652 fprintf(fout, "\n cond_c = (%s != 0);", buf1);
7653 pfomask &= ~PFOB_C;
7654 }
7655 output_std_flags(fout, po, &pfomask, buf1);
7656 break;
7657
7658 case OP_IMUL:
7659 if (po->operand_cnt == 2) {
7660 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7661 goto dualop_arith;
7662 }
7663 if (po->operand_cnt == 3)
7664 ferr(po, "TODO imul3\n");
7665 // fallthrough
7666 case OP_MUL:
7667 assert_operand_cnt(1);
7668 switch (po->operand[0].lmod) {
7669 case OPLM_DWORD:
7670 strcpy(buf1, po->op == OP_IMUL ? "(s64)(s32)" : "(u64)");
7671 fprintf(fout, " tmp64 = %seax * %s%s;\n", buf1, buf1,
7672 out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[0]));
7673 fprintf(fout, " edx = tmp64 >> 32;\n");
7674 fprintf(fout, " eax = tmp64;");
7675 break;
7676 case OPLM_BYTE:
7677 strcpy(buf1, po->op == OP_IMUL ? "(s16)(s8)" : "(u16)(u8)");
7678 fprintf(fout, " LOWORD(eax) = %seax * %s;", buf1,
7679 out_src_opr(buf2, sizeof(buf2), po, &po->operand[0],
7680 buf1, 0));
7681 break;
7682 default:
7683 ferr(po, "TODO: unhandled mul type\n");
7684 break;
7685 }
7686 last_arith_dst = NULL;
7687 delayed_flag_op = NULL;
7688 break;
7689
7690 case OP_DIV:
7691 case OP_IDIV:
7692 assert_operand_cnt(1);
7693 out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
7694 strcpy(cast, lmod_cast(po, po->operand[0].lmod,
7695 po->op == OP_IDIV));
7696 switch (po->operand[0].lmod) {
7697 case OPLM_DWORD:
7698 if (po->flags & OPF_32BIT)
7699 snprintf(buf2, sizeof(buf2), "%seax", cast);
7700 else {
7701 fprintf(fout, " tmp64 = ((u64)edx << 32) | eax;\n");
7702 snprintf(buf2, sizeof(buf2), "%stmp64",
7703 (po->op == OP_IDIV) ? "(s64)" : "");
7704 }
7705 if (po->operand[0].type == OPT_REG
7706 && po->operand[0].reg == xDX)
7707 {
7708 fprintf(fout, " eax = %s / %s%s;\n", buf2, cast, buf1);
7709 fprintf(fout, " edx = %s %% %s%s;", buf2, cast, buf1);
7710 }
7711 else {
7712 fprintf(fout, " edx = %s %% %s%s;\n", buf2, cast, buf1);
7713 fprintf(fout, " eax = %s / %s%s;", buf2, cast, buf1);
7714 }
7715 break;
7716 case OPLM_WORD:
7717 fprintf(fout, " tmp = (edx << 16) | (eax & 0xffff);\n");
7718 snprintf(buf2, sizeof(buf2), "%stmp",
7719 (po->op == OP_IDIV) ? "(s32)" : "");
7720 if (po->operand[0].type == OPT_REG
7721 && po->operand[0].reg == xDX)
7722 {
7723 fprintf(fout, " LOWORD(eax) = %s / %s%s;\n",
7724 buf2, cast, buf1);
7725 fprintf(fout, " LOWORD(edx) = %s %% %s%s;",
7726 buf2, cast, buf1);
7727 }
7728 else {
7729 fprintf(fout, " LOWORD(edx) = %s %% %s%s;\n",
7730 buf2, cast, buf1);
7731 fprintf(fout, " LOWORD(eax) = %s / %s%s;",
7732 buf2, cast, buf1);
7733 }
7734 strcat(g_comment, " div16");
7735 break;
7736 default:
7737 ferr(po, "unhandled div lmod %d\n", po->operand[0].lmod);
7738 }
7739 last_arith_dst = NULL;
7740 delayed_flag_op = NULL;
7741 break;
7742
7743 case OP_TEST:
7744 case OP_CMP:
7745 propagate_lmod(po, &po->operand[0], &po->operand[1]);
7746 if (pfomask != 0) {
7747 for (j = 0; j < 8; j++) {
7748 if (pfomask & (1 << j)) {
7749 out_cmp_test(buf1, sizeof(buf1), po, j, 0);
7750 fprintf(fout, " cond_%s = %s;",
7751 parsed_flag_op_names[j], buf1);
7752 }
7753 }
7754 pfomask = 0;
7755 }
7756 else
7757 no_output = 1;
7758 last_arith_dst = NULL;
7759 delayed_flag_op = po;
7760 break;
7761
7762 case OP_SCC:
7763 // SETcc - should already be handled
7764 break;
7765
7766 // note: we reuse OP_Jcc for SETcc, only flags differ
7767 case OP_JCC:
7768 fprintf(fout, "\n goto %s;", po->operand[0].name);
7769 break;
7770
7771 case OP_JECXZ:
7772 fprintf(fout, " if (ecx == 0)\n");
7773 fprintf(fout, " goto %s;", po->operand[0].name);
7774 strcat(g_comment, " jecxz");
7775 break;
7776
7777 case OP_LOOP:
7778 fprintf(fout, " if (--ecx != 0)\n");
7779 fprintf(fout, " goto %s;", po->operand[0].name);
7780 strcat(g_comment, " loop");
7781 break;
7782
7783 case OP_JMP:
7784 assert_operand_cnt(1);
7785 last_arith_dst = NULL;
7786 delayed_flag_op = NULL;
7787
7788 if (po->operand[0].type == OPT_REGMEM) {
7789 ret = sscanf(po->operand[0].name, "%[^[][%[^*]*4]",
7790 buf1, buf2);
7791 if (ret != 2)
7792 ferr(po, "parse failure for jmp '%s'\n",
7793 po->operand[0].name);
7794 fprintf(fout, " goto *jt_%s[%s];", buf1, buf2);
7795 break;
7796 }
7797 else if (po->operand[0].type != OPT_LABEL)
7798 ferr(po, "unhandled jmp type\n");
7799
7800 fprintf(fout, " goto %s;", po->operand[0].name);
7801 break;
7802
7803 case OP_CALL:
7804 assert_operand_cnt(1);
7805 pp = po->pp;
7806 my_assert_not(pp, NULL);
7807
7808 strcpy(buf3, " ");
7809 if (po->flags & OPF_CC) {
7810 // we treat conditional branch to another func
7811 // (yes such code exists..) as conditional tailcall
7812 strcat(buf3, " ");
7813 fprintf(fout, " {\n");
7814 }
7815
7816 if (pp->is_fptr && !pp->is_arg) {
7817 fprintf(fout, "%s%s = %s;\n", buf3, pp->name,
7818 out_src_opr(buf1, sizeof(buf1), po, &po->operand[0],
7819 "(void *)", 0));
7820 }
7821 if (pp->is_fptr && (pp->is_unresolved || pp->is_guessed)) {
7822 fprintf(fout, "%sunresolved_call(\"%s:%d\", %s);\n",
7823 buf3, asmfn, po->asmln, pp->name);
7824 }
7825
7826 fprintf(fout, "%s", buf3);
7827 if (strstr(pp->ret_type.name, "int64")) {
7828 if (po->flags & OPF_TAIL)
7829 ferr(po, "int64 and tail?\n");
7830 fprintf(fout, "tmp64 = ");
7831 }
7832 else if (!IS(pp->ret_type.name, "void")) {
7833 if (po->flags & OPF_TAIL) {
7834 if (regmask_ret & mxAX) {
7835 fprintf(fout, "return ");
7836 if (g_func_pp->ret_type.is_ptr != pp->ret_type.is_ptr)
7837 fprintf(fout, "(%s)", g_func_pp->ret_type.name);
7838 }
7839 else if (regmask_ret & mxST0)
7840 ferr(po, "float tailcall\n");
7841 }
7842 else if (po->regmask_dst & mxAX) {
7843 fprintf(fout, "eax = ");
7844 if (pp->ret_type.is_ptr)
7845 fprintf(fout, "(u32)");
7846 }
7847 else if (po->regmask_dst & mxST0) {
7848 ferr_assert(po, po->flags & OPF_FPUSH);
7849 if (need_float_stack)
7850 fprintf(fout, "f_st[--f_stp & 7] = ");
7851 else
7852 fprintf(fout, "f_st0 = ");
7853 }
7854 }
7855
7856 if (pp->name[0] == 0)
7857 ferr(po, "missing pp->name\n");
7858 fprintf(fout, "%s%s(", pp->name,
7859 pp->has_structarg ? "_sa" : "");
7860
7861 if (po->flags & OPF_ATAIL) {
7862 int check_compat =
7863 g_func_pp->is_stdcall && g_func_pp->argc_stack > 0;
7864 check_compat |= pp->argc_stack > 0;
7865 if (check_compat
7866 && (pp->argc_stack != g_func_pp->argc_stack
7867 || pp->is_stdcall != g_func_pp->is_stdcall))
7868 ferr(po, "incompatible arg-reuse tailcall\n");
7869 if (g_func_pp->has_retreg)
7870 ferr(po, "TODO: retreg+tailcall\n");
7871
7872 for (arg = j = 0; arg < pp->argc; arg++) {
7873 if (arg > 0)
7874 fprintf(fout, ", ");
7875
7876 cast[0] = 0;
7877 if (pp->arg[arg].type.is_ptr)
7878 snprintf(cast, sizeof(cast), "(%s)",
7879 pp->arg[arg].type.name);
7880
7881 if (pp->arg[arg].reg != NULL) {
7882 fprintf(fout, "%s%s", cast, pp->arg[arg].reg);
7883 continue;
7884 }
7885 // stack arg
7886 for (; j < g_func_pp->argc; j++)
7887 if (g_func_pp->arg[j].reg == NULL)
7888 break;
7889 fprintf(fout, "%sa%d", cast, j + 1);
7890 j++;
7891 }
7892 }
7893 else {
7894 for (arg = 0; arg < pp->argc; arg++) {
7895 if (arg > 0)
7896 fprintf(fout, ", ");
7897
7898 cast[0] = 0;
7899 if (pp->arg[arg].type.is_ptr)
7900 snprintf(cast, sizeof(cast), "(%s)",
7901 pp->arg[arg].type.name);
7902
7903 if (pp->arg[arg].reg != NULL) {
7904 if (pp->arg[arg].type.is_retreg)
7905 fprintf(fout, "&%s", pp->arg[arg].reg);
7906 else if (IS(pp->arg[arg].reg, "ebp")
7907 && g_bp_frame && !(po->flags & OPF_EBP_S))
7908 {
7909 // rare special case
7910 fprintf(fout, "%s(u32)&sf.b[sizeof(sf)]", cast);
7911 strcat(g_comment, " bp_ref");
7912 }
7913 else
7914 fprintf(fout, "%s%s", cast, pp->arg[arg].reg);
7915 continue;
7916 }
7917
7918 // stack arg
7919 if (pp->arg[arg].push_ref_cnt == 0)
7920 ferr(po, "parsed_op missing for arg%d\n", arg);
7921 if (pp->arg[arg].push_ref_cnt > 1)
7922 ferr_assert(po, pp->arg[arg].is_saved);
7923 tmp_op = pp->arg[arg].push_refs[0];
7924 ferr_assert(po, tmp_op != NULL);
7925
7926 if (tmp_op->flags & OPF_VAPUSH) {
7927 fprintf(fout, "ap");
7928 }
7929 else if (tmp_op->op == OP_FST) {
7930 fprintf(fout, "fs_%d", tmp_op->p_argnum);
7931 if (tmp_op->operand[0].lmod == OPLM_QWORD)
7932 arg++;
7933 }
7934 else if (pp->arg[arg].type.is_64bit) {
7935 ferr_assert(po, tmp_op->p_argpass == 0);
7936 ferr_assert(po, !pp->arg[arg].is_saved);
7937 ferr_assert(po, !pp->arg[arg].type.is_float);
7938 ferr_assert(po, cast[0] == 0);
7939 out_src_opr(buf1, sizeof(buf1),
7940 tmp_op, &tmp_op->operand[0], cast, 0);
7941 arg++;
7942 ferr_assert(po, pp->arg[arg].push_ref_cnt == 1);
7943 tmp_op = pp->arg[arg].push_refs[0];
7944 ferr_assert(po, tmp_op != NULL);
7945 out_src_opr(buf2, sizeof(buf2),
7946 tmp_op, &tmp_op->operand[0], cast, 0);
7947 fprintf(fout, "((u64)(%s) << 32) | (%s)",
7948 buf2, buf1);
7949 }
7950 else if (tmp_op->p_argpass != 0) {
7951 ferr_assert(po, !pp->arg[arg].type.is_float);
7952 fprintf(fout, "a%d", tmp_op->p_argpass);
7953 }
7954 else if (pp->arg[arg].is_saved) {
7955 ferr_assert(po, tmp_op->p_argnum > 0);
7956 ferr_assert(po, !pp->arg[arg].type.is_float);
7957 fprintf(fout, "%s%s", cast,
7958 saved_arg_name(buf1, sizeof(buf1),
7959 tmp_op->p_arggrp, tmp_op->p_argnum));
7960 }
7961 else if (pp->arg[arg].type.is_float) {
7962 ferr_assert(po, !pp->arg[arg].type.is_64bit);
7963 fprintf(fout, "%s",
7964 out_src_opr_float(buf1, sizeof(buf1),
7965 tmp_op, &tmp_op->operand[0], need_float_stack));
7966 }
7967 else {
7968 fprintf(fout, "%s",
7969 out_src_opr(buf1, sizeof(buf1),
7970 tmp_op, &tmp_op->operand[0], cast, 0));
7971 }
7972 }
7973 }
7974 fprintf(fout, ");");
7975
7976 if (strstr(pp->ret_type.name, "int64")) {
7977 fprintf(fout, "\n");
7978 fprintf(fout, "%sedx = tmp64 >> 32;\n", buf3);
7979 fprintf(fout, "%seax = tmp64;", buf3);
7980 }
7981
7982 if (pp->is_unresolved) {
7983 snprintf(buf2, sizeof(buf2), " unresolved %dreg",
7984 pp->argc_reg);
7985 strcat(g_comment, buf2);
7986 }
7987
7988 if (po->flags & OPF_TAIL) {
7989 ret = 0;
7990 if (i == opcnt - 1 || pp->is_noreturn)
7991 ret = 0;
7992 else if (IS(pp->ret_type.name, "void"))
7993 ret = 1;
7994 else if (!(regmask_ret & (1 << xAX)))
7995 ret = 1;
7996 // else already handled as 'return f()'
7997
7998 if (ret) {
7999 fprintf(fout, "\n%sreturn;", buf3);
8000 strcat(g_comment, " ^ tailcall");
8001 }
8002 else
8003 strcat(g_comment, " tailcall");
8004
8005 if ((regmask_ret & (1 << xAX))
8006 && IS(pp->ret_type.name, "void") && !pp->is_noreturn)
8007 {
8008 ferr(po, "int func -> void func tailcall?\n");
8009 }
8010 }
8011 if (pp->is_noreturn)
8012 strcat(g_comment, " noreturn");
8013 if ((po->flags & OPF_ATAIL) && pp->argc_stack > 0)
8014 strcat(g_comment, " argframe");
8015 if (po->flags & OPF_CC)
8016 strcat(g_comment, " cond");
8017
8018 if (po->flags & OPF_CC)
8019 fprintf(fout, "\n }");
8020
8021 delayed_flag_op = NULL;
8022 last_arith_dst = NULL;
8023 break;
8024
8025 case OP_RET:
8026 do_tail:
8027 if (g_func_pp->is_vararg)
8028 fprintf(fout, " va_end(ap);\n");
8029 if (g_func_pp->has_retreg) {
8030 for (arg = 0; arg < g_func_pp->argc; arg++)
8031 if (g_func_pp->arg[arg].type.is_retreg)
8032 fprintf(fout, " *r_%s = %s;\n",
8033 g_func_pp->arg[arg].reg, g_func_pp->arg[arg].reg);
8034 }
8035
8036 if (regmask_ret & mxST0) {
8037 fprintf(fout, " return %s;", float_st0);
8038 }
8039 else if (!(regmask_ret & mxAX)) {
8040 if (i != opcnt - 1 || label_pending)
8041 fprintf(fout, " return;");
8042 }
8043 else if (g_func_pp->ret_type.is_ptr) {
8044 fprintf(fout, " return (%s)eax;",
8045 g_func_pp->ret_type.name);
8046 }
8047 else if (IS(g_func_pp->ret_type.name, "__int64"))
8048 fprintf(fout, " return ((u64)edx << 32) | eax;");
8049 else
8050 fprintf(fout, " return eax;");
8051
8052 last_arith_dst = NULL;
8053 delayed_flag_op = NULL;
8054 break;
8055
8056 case OP_PUSH:
8057 out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
8058 if (po->p_argnum != 0) {
8059 // special case - saved func arg
8060 fprintf(fout, " %s = %s;",
8061 saved_arg_name(buf2, sizeof(buf2),
8062 po->p_arggrp, po->p_argnum), buf1);
8063 break;
8064 }
8065 else if (po->flags & OPF_RSAVE) {
8066 fprintf(fout, " s_%s = %s;", buf1, buf1);
8067 break;
8068 }
8069 else if (po->flags & OPF_PPUSH) {
8070 tmp_op = po->datap;
8071 ferr_assert(po, tmp_op != NULL);
8072 out_dst_opr(buf2, sizeof(buf2), po, &tmp_op->operand[0]);
8073 fprintf(fout, " pp_%s = %s;", buf2, buf1);
8074 break;
8075 }
8076 else if (g_func_pp->is_userstack) {
8077 fprintf(fout, " *(--esp) = %s;", buf1);
8078 break;
8079 }
8080 if (!(g_ida_func_attr & IDAFA_NORETURN))
8081 ferr(po, "stray push encountered\n");
8082 no_output = 1;
8083 break;
8084
8085 case OP_POP:
8086 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
8087 if (po->flags & OPF_RSAVE) {
8088 fprintf(fout, " %s = s_%s;", buf1, buf1);
8089 break;
8090 }
8091 else if (po->flags & OPF_PPUSH) {
8092 // push/pop graph / non-const
8093 ferr_assert(po, po->datap == NULL);
8094 fprintf(fout, " %s = pp_%s;", buf1, buf1);
8095 break;
8096 }
8097 else if (po->datap != NULL) {
8098 // push/pop pair
8099 tmp_op = po->datap;
8100 fprintf(fout, " %s = %s;", buf1,
8101 out_src_opr(buf2, sizeof(buf2),
8102 tmp_op, &tmp_op->operand[0],
8103 default_cast_to(buf3, sizeof(buf3), &po->operand[0]), 0));
8104 break;
8105 }
8106 else if (g_func_pp->is_userstack) {
8107 fprintf(fout, " %s = *esp++;", buf1);
8108 break;
8109 }
8110 else
8111 ferr(po, "stray pop encountered\n");
8112 break;
8113
8114 case OP_NOP:
8115 no_output = 1;
8116 break;
8117
8118 // pseudo ops
8119 case OPP_ALLSHL:
8120 case OPP_ALLSHR:
8121 fprintf(fout, " tmp64 = ((u64)edx << 32) | eax;\n");
8122 fprintf(fout, " tmp64 = (s64)tmp64 %s LOBYTE(ecx);\n",
8123 po->op == OPP_ALLSHL ? "<<" : ">>");
8124 fprintf(fout, " edx = tmp64 >> 32; eax = tmp64;");
8125 strcat(g_comment, po->op == OPP_ALLSHL
8126 ? " allshl" : " allshr");
8127 break;
8128
8129 // x87
8130 case OP_FLD:
8131 if (need_float_stack) {
8132 out_src_opr_float(buf1, sizeof(buf1),
8133 po, &po->operand[0], 1);
8134 if (po->regmask_src & mxSTa) {
8135 fprintf(fout, " f_st[(f_stp - 1) & 7] = %s; f_stp--;",
8136 buf1);
8137 }
8138 else
8139 fprintf(fout, " f_st[--f_stp & 7] = %s;", buf1);
8140 }
8141 else {
8142 if (po->flags & OPF_FSHIFT)
8143 fprintf(fout, " f_st1 = f_st0;");
8144 if (po->operand[0].type == OPT_REG
8145 && po->operand[0].reg == xST0)
8146 {
8147 strcat(g_comment, " fld st");
8148 break;
8149 }
8150 fprintf(fout, " f_st0 = %s;",
8151 out_src_opr_float(buf1, sizeof(buf1),
8152 po, &po->operand[0], 0));
8153 }
8154 strcat(g_comment, " fld");
8155 break;
8156
8157 case OP_FILD:
8158 out_src_opr(buf1, sizeof(buf1), po, &po->operand[0],
8159 lmod_cast(po, po->operand[0].lmod, 1), 0);
8160 snprintf(buf2, sizeof(buf2), "(%s)%s", float_type, buf1);
8161 if (need_float_stack) {
8162 fprintf(fout, " f_st[--f_stp & 7] = %s;", buf2);
8163 }
8164 else {
8165 if (po->flags & OPF_FSHIFT)
8166 fprintf(fout, " f_st1 = f_st0;");
8167 fprintf(fout, " f_st0 = %s;", buf2);
8168 }
8169 strcat(g_comment, " fild");
8170 break;
8171
8172 case OP_FLDc:
8173 if (need_float_stack)
8174 fprintf(fout, " f_st[--f_stp & 7] = ");
8175 else {
8176 if (po->flags & OPF_FSHIFT)
8177 fprintf(fout, " f_st1 = f_st0;");
8178 fprintf(fout, " f_st0 = ");
8179 }
8180 switch (po->operand[0].val) {
8181 case X87_CONST_1: fprintf(fout, "1.0;"); break;
8182 case X87_CONST_L2T: fprintf(fout, "3.321928094887362;"); break;
8183 case X87_CONST_L2E: fprintf(fout, "M_LOG2E;"); break;
8184 case X87_CONST_PI: fprintf(fout, "M_PI;"); break;
8185 case X87_CONST_LG2: fprintf(fout, "0.301029995663981;"); break;
8186 case X87_CONST_LN2: fprintf(fout, "M_LN2;"); break;
8187 case X87_CONST_Z: fprintf(fout, "0.0;"); break;
8188 default: ferr_assert(po, 0); break;
8189 }
8190 break;
8191
8192 case OP_FST:
8193 dead_dst = 0;
8194 if (po->flags & OPF_FARG) {
8195 // store to stack as func arg
8196 fprintf(fout, " fs_%d = %s;", po->p_argnum, float_st0);
8197 }
8198 else if (po->operand[0].type == OPT_REG
8199 && po->operand[0].reg == xST0)
8200 {
8201 dead_dst = 1;
8202 }
8203 else if (float_opr_needs_helper(po, &po->operand[0])) {
8204 out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], "", 1);
8205 fprintf(fout, " %s_store(%s, %s);",
8206 po->operand[0].lmod == OPLM_QWORD ? "double" : "float",
8207 float_st0, buf1);
8208 }
8209 else {
8210 out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
8211 need_float_stack);
8212 fprintf(fout, " %s = %s;", buf1, float_st0);
8213 }
8214 if (po->flags & OPF_FSHIFT) {
8215 if (need_float_stack)
8216 fprintf(fout, " f_stp++;");
8217 else
8218 fprintf(fout, " f_st0 = f_st1;");
8219 }
8220 if (dead_dst && !(po->flags & OPF_FSHIFT))
8221 no_output = 1;
8222 else
8223 strcat(g_comment, " fst");
8224 break;
8225
8226 case OP_FIST:
8227 fprintf(fout, " %s = %s%s;",
8228 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
8229 lmod_cast(po, po->operand[0].lmod, 1), float_st0);
8230 if (po->flags & OPF_FSHIFT) {
8231 if (need_float_stack)
8232 fprintf(fout, " f_stp++;");
8233 else
8234 fprintf(fout, " f_st0 = f_st1;");
8235 }
8236 strcat(g_comment, " fist");
8237 break;
8238
8239 case OP_FABS:
8240 fprintf(fout, " %s = fabs%s(%s);", float_st0,
8241 need_double ? "" : "f", float_st0);
8242 break;
8243
8244 case OP_FADD:
8245 case OP_FDIV:
8246 case OP_FMUL:
8247 case OP_FSUB:
8248 out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
8249 need_float_stack);
8250 out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1],
8251 need_float_stack);
8252 dead_dst = (po->flags & OPF_FPOP)
8253 && po->operand[0].type == OPT_REG
8254 && po->operand[0].reg == xST0;
8255 switch (po->op) {
8256 case OP_FADD: j = '+'; break;
8257 case OP_FDIV: j = '/'; break;
8258 case OP_FMUL: j = '*'; break;
8259 case OP_FSUB: j = '-'; break;
8260 default: j = 'x'; break;
8261 }
8262 if (need_float_stack) {
8263 if (!dead_dst)
8264 fprintf(fout, " %s %c= %s;", buf1, j, buf2);
8265 if (po->flags & OPF_FSHIFT)
8266 fprintf(fout, " f_stp++;");
8267 }
8268 else {
8269 if (po->flags & OPF_FSHIFT) {
8270 // note: assumes only 2 regs handled
8271 if (!dead_dst)
8272 fprintf(fout, " f_st0 = f_st1 %c f_st0;", j);
8273 else
8274 fprintf(fout, " f_st0 = f_st1;");
8275 }
8276 else if (!dead_dst)
8277 fprintf(fout, " %s %c= %s;", buf1, j, buf2);
8278 }
8279 no_output = (dead_dst && !(po->flags & OPF_FSHIFT));
8280 break;
8281
8282 case OP_FDIVR:
8283 case OP_FSUBR:
8284 out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
8285 need_float_stack);
8286 out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1],
8287 need_float_stack);
8288 out_src_opr_float(buf3, sizeof(buf3), po, &po->operand[0],
8289 need_float_stack);
8290 dead_dst = (po->flags & OPF_FPOP)
8291 && po->operand[0].type == OPT_REG
8292 && po->operand[0].reg == xST0;
8293 j = po->op == OP_FDIVR ? '/' : '-';
8294 if (need_float_stack) {
8295 if (!dead_dst)
8296 fprintf(fout, " %s = %s %c %s;", buf1, buf2, j, buf3);
8297 if (po->flags & OPF_FSHIFT)
8298 fprintf(fout, " f_stp++;");
8299 }
8300 else {
8301 if (po->flags & OPF_FSHIFT) {
8302 if (!dead_dst)
8303 fprintf(fout, " f_st0 = f_st0 %c f_st1;", j);
8304 else
8305 fprintf(fout, " f_st0 = f_st1;");
8306 }
8307 else if (!dead_dst)
8308 fprintf(fout, " %s = %s %c %s;", buf1, buf2, j, buf3);
8309 }
8310 no_output = (dead_dst && !(po->flags & OPF_FSHIFT));
8311 break;
8312
8313 case OP_FIADD:
8314 case OP_FIDIV:
8315 case OP_FIMUL:
8316 case OP_FISUB:
8317 switch (po->op) {
8318 case OP_FIADD: j = '+'; break;
8319 case OP_FIDIV: j = '/'; break;
8320 case OP_FIMUL: j = '*'; break;
8321 case OP_FISUB: j = '-'; break;
8322 default: j = 'x'; break;
8323 }
8324 fprintf(fout, " %s %c= (%s)%s;", float_st0,
8325 j, float_type,
8326 out_src_opr(buf1, sizeof(buf1), po, &po->operand[0],
8327 lmod_cast(po, po->operand[0].lmod, 1), 0));
8328 break;
8329
8330 case OP_FIDIVR:
8331 case OP_FISUBR:
8332 fprintf(fout, " %s = %s %c %s;", float_st0,
8333 out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1],
8334 need_float_stack),
8335 po->op == OP_FIDIVR ? '/' : '-', float_st0);
8336 break;
8337
8338 case OP_FCOM: {
8339 int mask, z_check;
8340 ferr_assert(po, po->datap != NULL);
8341 mask = (long)po->datap & 0xffff;
8342 z_check = ((long)po->datap >> 16) & 1;
8343 out_src_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
8344 need_float_stack);
8345 if (mask == 0x0100 || mask == 0x0500) { // C0 -> <
8346 fprintf(fout, " f_sw = %s < %s ? 0x0100 : 0;",
8347 float_st0, buf1);
8348 }
8349 else if (mask == 0x4000 || mask == 0x4400) { // C3 -> =
8350 fprintf(fout, " f_sw = %s == %s ? 0x4000 : 0;",
8351 float_st0, buf1);
8352 }
8353 else if (mask == 0x4100) { // C3, C0
8354 if (z_check) {
8355 fprintf(fout, " f_sw = %s <= %s ? 0x4100 : 0;",
8356 float_st0, buf1);
8357 strcat(g_comment, " z_chk_det");
8358 }
8359 else {
8360 fprintf(fout, " f_sw = %s == %s ? 0x4000 : "
8361 "(%s < %s ? 0x0100 : 0);",
8362 float_st0, buf1, float_st0, buf1);
8363 }
8364 }
8365 else
8366 ferr(po, "unhandled sw mask: %x\n", mask);
8367 if (po->flags & OPF_FSHIFT) {
8368 if (need_float_stack) {
8369 if (po->flags & OPF_FPOPP)
8370 fprintf(fout, " f_stp += 2;");
8371 else
8372 fprintf(fout, " f_stp++;");
8373 }
8374 else {
8375 ferr_assert(po, !(po->flags & OPF_FPOPP));
8376 fprintf(fout, " f_st0 = f_st1;");
8377 }
8378 }
8379 break;
8380 }
8381
8382 case OP_FNSTSW:
8383 fprintf(fout, " %s = f_sw;",
8384 out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]));
8385 break;
8386
8387 case OP_FCHS:
8388 fprintf(fout, " %s = -%s;", float_st0, float_st0);
8389 break;
8390
8391 case OP_FCOS:
8392 fprintf(fout, " %s = cos%s(%s);", float_st0,
8393 need_double ? "" : "f", float_st0);
8394 break;
8395
8396 case OP_FPATAN:
8397 if (need_float_stack) {
8398 fprintf(fout, " %s = atan%s(%s / %s);", float_st1,
8399 need_double ? "" : "f", float_st1, float_st0);
8400 fprintf(fout, " f_stp++;");
8401 }
8402 else {
8403 fprintf(fout, " f_st0 = atan%s(f_st1 / f_st0);",
8404 need_double ? "" : "f");
8405 }
8406 break;
8407
8408 case OP_FYL2X:
8409 if (need_float_stack) {
8410 fprintf(fout, " %s = %s * log2%s(%s);", float_st1,
8411 float_st1, need_double ? "" : "f", float_st0);
8412 fprintf(fout, " f_stp++;");
8413 }
8414 else {
8415 fprintf(fout, " f_st0 = f_st1 * log2%s(f_st0);",
8416 need_double ? "" : "f");
8417 }
8418 strcat(g_comment, " fyl2x");
8419 break;
8420
8421 case OP_FSIN:
8422 fprintf(fout, " %s = sin%s(%s);", float_st0,
8423 need_double ? "" : "f", float_st0);
8424 break;
8425
8426 case OP_FSQRT:
8427 fprintf(fout, " %s = sqrt%s(%s);", float_st0,
8428 need_double ? "" : "f", float_st0);
8429 break;
8430
8431 case OP_FXCH:
8432 dead_dst = po->operand[0].type == OPT_REG
8433 && po->operand[0].reg == xST0;
8434 if (!dead_dst) {
8435 out_src_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
8436 need_float_stack);
8437 fprintf(fout, " { %s t = %s; %s = %s; %s = t; }", float_type,
8438 float_st0, float_st0, buf1, buf1);
8439 strcat(g_comment, " fxch");
8440 }
8441 else
8442 no_output = 1;
8443 break;
8444
8445 case OPP_FTOL:
8446 ferr_assert(po, po->flags & OPF_32BIT);
8447 fprintf(fout, " eax = (s32)%s;", float_st0);
8448 if (po->flags & OPF_FSHIFT) {
8449 if (need_float_stack)
8450 fprintf(fout, " f_stp++;");
8451 else
8452 fprintf(fout, " f_st0 = f_st1;");
8453 }
8454 strcat(g_comment, " ftol");
8455 goto tail_check;
8456
8457 case OPP_CIPOW:
8458 if (need_float_stack) {
8459 fprintf(fout, " %s = pow%s(%s, %s);", float_st1,
8460 need_double ? "" : "f", float_st1, float_st0);
8461 fprintf(fout, " f_stp++;");
8462 }
8463 else {
8464 fprintf(fout, " f_st0 = pow%s(f_st1, f_st0);",
8465 need_double ? "" : "f");
8466 }
8467 strcat(g_comment, " CIpow");
8468 goto tail_check;
8469
8470 case OPP_ABORT:
8471 fprintf(fout, " do_skip_code_abort();");
8472 break;
8473
8474 // mmx
8475 case OP_EMMS:
8476 fprintf(fout, " do_emms();");
8477 break;
8478
8479 tail_check:
8480 if (po->flags & OPF_TAIL) {
8481 fprintf(fout, "\n");
8482 strcat(g_comment, " tail");
8483 goto do_tail;
8484 }
8485 break;
8486
8487 default:
8488 no_output = 1;
8489 ferr(po, "unhandled op type %d, flags %x\n",
8490 po->op, po->flags);
8491 break;
8492 }
8493
8494 if (g_comment[0] != 0) {
8495 char *p = g_comment;
8496 while (my_isblank(*p))
8497 p++;
8498 fprintf(fout, " // %s", p);
8499 g_comment[0] = 0;
8500 no_output = 0;
8501 }
8502 if (!no_output)
8503 fprintf(fout, "\n");
8504
8505 // some sanity checking
8506 if (po->flags & OPF_REP) {
8507 if (po->op != OP_STOS && po->op != OP_MOVS
8508 && po->op != OP_CMPS && po->op != OP_SCAS)
8509 ferr(po, "unexpected rep\n");
8510 if (!(po->flags & (OPF_REPZ|OPF_REPNZ))
8511 && (po->op == OP_CMPS || po->op == OP_SCAS))
8512 ferr(po, "cmps/scas with plain rep\n");
8513 }
8514 if ((po->flags & (OPF_REPZ|OPF_REPNZ))
8515 && po->op != OP_CMPS && po->op != OP_SCAS)
8516 ferr(po, "unexpected repz/repnz\n");
8517
8518 if (pfomask != 0)
8519 ferr(po, "missed flag calc, pfomask=%x\n", pfomask);
8520
8521 if ((po->flags & OPF_LOCK) && !lock_handled)
8522 ferr(po, "unhandled lock\n");
8523
8524 // see is delayed flag stuff is still valid
8525 if (delayed_flag_op != NULL && delayed_flag_op != po) {
8526 if (is_any_opr_modified(delayed_flag_op, po, 0))
8527 delayed_flag_op = NULL;
8528 }
8529
8530 if (last_arith_dst != NULL && last_arith_dst != &po->operand[0]) {
8531 if (is_opr_modified(last_arith_dst, po))
8532 last_arith_dst = NULL;
8533 }
8534
8535 if (!no_output)
8536 label_pending = 0;
8537 }
8538
8539 if (g_stack_fsz && !g_stack_frame_used)
8540 fprintf(fout, " (void)sf;\n");
8541
8542 fprintf(fout, "}\n\n");
8543
8544 gen_x_cleanup(opcnt);
8545}
8546
8547static void gen_x_cleanup(int opcnt)
8548{
8549 int i;
8550
8551 for (i = 0; i < opcnt; i++) {
8552 struct label_ref *lr, *lr_del;
8553
8554 lr = g_label_refs[i].next;
8555 while (lr != NULL) {
8556 lr_del = lr;
8557 lr = lr->next;
8558 free(lr_del);
8559 }
8560 g_label_refs[i].i = -1;
8561 g_label_refs[i].next = NULL;
8562
8563 if (ops[i].op == OP_CALL) {
8564 if (ops[i].pp)
8565 proto_release(ops[i].pp);
8566 }
8567 }
8568 g_func_pp = NULL;
8569}
8570
8571struct func_proto_dep;
8572
8573struct func_prototype {
8574 char name[NAMELEN];
8575 int id;
8576 int argc_stack;
8577 int regmask_dep; // likely register args
8578 int regmask_use; // used registers
8579 int has_ret:3; // -1, 0, 1: unresolved, no, yes
8580 unsigned int has_ret64:1;
8581 unsigned int dep_resolved:1;
8582 unsigned int is_stdcall:1;
8583 unsigned int eax_pass:1; // returns without touching eax
8584 unsigned int ptr_taken:1; // pointer taken of this func
8585 struct func_proto_dep *dep_func;
8586 int dep_func_cnt;
8587 const struct parsed_proto *pp; // seed pp, if any
8588};
8589
8590struct func_proto_dep {
8591 char *name;
8592 struct func_prototype *proto;
8593 int regmask_live; // .. at the time of call
8594 unsigned int ret_dep:1; // return from this is caller's return
8595 unsigned int has_ret:1; // found from eax use after return
8596 unsigned int has_ret64:1;
8597 unsigned int ptr_taken:1; // pointer taken, not a call
8598};
8599
8600static struct func_prototype *hg_fp;
8601static int hg_fp_cnt;
8602
8603static struct scanned_var {
8604 char name[NAMELEN];
8605 enum opr_lenmod lmod;
8606 unsigned int is_seeded:1;
8607 unsigned int is_c_str:1;
8608 const struct parsed_proto *pp; // seed pp, if any
8609} *hg_vars;
8610static int hg_var_cnt;
8611
8612static char **hg_refs;
8613static int hg_ref_cnt;
8614
8615static void output_hdr_fp(FILE *fout, const struct func_prototype *fp,
8616 int count);
8617
8618static struct func_prototype *hg_fp_add(const char *funcn)
8619{
8620 struct func_prototype *fp;
8621
8622 if ((hg_fp_cnt & 0xff) == 0) {
8623 hg_fp = realloc(hg_fp, sizeof(hg_fp[0]) * (hg_fp_cnt + 0x100));
8624 my_assert_not(hg_fp, NULL);
8625 memset(hg_fp + hg_fp_cnt, 0, sizeof(hg_fp[0]) * 0x100);
8626 }
8627
8628 fp = &hg_fp[hg_fp_cnt];
8629 snprintf(fp->name, sizeof(fp->name), "%s", funcn);
8630 fp->id = hg_fp_cnt;
8631 fp->argc_stack = -1;
8632 hg_fp_cnt++;
8633
8634 return fp;
8635}
8636
8637static struct func_proto_dep *hg_fp_find_dep(struct func_prototype *fp,
8638 const char *name)
8639{
8640 int i;
8641
8642 for (i = 0; i < fp->dep_func_cnt; i++)
8643 if (IS(fp->dep_func[i].name, name))
8644 return &fp->dep_func[i];
8645
8646 return NULL;
8647}
8648
8649static void hg_fp_add_dep(struct func_prototype *fp, const char *name,
8650 unsigned int ptr_taken)
8651{
8652 struct func_proto_dep * dep;
8653
8654 // is it a dupe?
8655 dep = hg_fp_find_dep(fp, name);
8656 if (dep != NULL && dep->ptr_taken == ptr_taken)
8657 return;
8658
8659 if ((fp->dep_func_cnt & 0xff) == 0) {
8660 fp->dep_func = realloc(fp->dep_func,
8661 sizeof(fp->dep_func[0]) * (fp->dep_func_cnt + 0x100));
8662 my_assert_not(fp->dep_func, NULL);
8663 memset(&fp->dep_func[fp->dep_func_cnt], 0,
8664 sizeof(fp->dep_func[0]) * 0x100);
8665 }
8666 fp->dep_func[fp->dep_func_cnt].name = strdup(name);
8667 fp->dep_func[fp->dep_func_cnt].ptr_taken = ptr_taken;
8668 fp->dep_func_cnt++;
8669}
8670
8671static int hg_fp_cmp_name(const void *p1_, const void *p2_)
8672{
8673 const struct func_prototype *p1 = p1_, *p2 = p2_;
8674 return strcmp(p1->name, p2->name);
8675}
8676
8677#if 0
8678static int hg_fp_cmp_id(const void *p1_, const void *p2_)
8679{
8680 const struct func_prototype *p1 = p1_, *p2 = p2_;
8681 return p1->id - p2->id;
8682}
8683#endif
8684
8685static void hg_ref_add(const char *name)
8686{
8687 if ((hg_ref_cnt & 0xff) == 0) {
8688 hg_refs = realloc(hg_refs, sizeof(hg_refs[0]) * (hg_ref_cnt + 0x100));
8689 my_assert_not(hg_refs, NULL);
8690 memset(hg_refs + hg_ref_cnt, 0, sizeof(hg_refs[0]) * 0x100);
8691 }
8692
8693 hg_refs[hg_ref_cnt] = strdup(name);
8694 my_assert_not(hg_refs[hg_ref_cnt], NULL);
8695 hg_ref_cnt++;
8696}
8697
8698// recursive register dep pass
8699// - track saved regs (part 2)
8700// - try to figure out arg-regs
8701// - calculate reg deps
8702static void gen_hdr_dep_pass(int i, int opcnt, unsigned char *cbits,
8703 struct func_prototype *fp, int regmask_save, int regmask_dst,
8704 int *regmask_dep, int *regmask_use, int *has_ret)
8705{
8706 struct func_proto_dep *dep;
8707 struct parsed_op *po;
8708 int from_caller = 0;
8709 int j, l;
8710 int reg;
8711 int ret;
8712
8713 for (; i < opcnt; i++)
8714 {
8715 if (cbits[i >> 3] & (1 << (i & 7)))
8716 return;
8717 cbits[i >> 3] |= (1 << (i & 7));
8718
8719 po = &ops[i];
8720
8721 if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
8722 if (po->flags & OPF_RMD)
8723 continue;
8724
8725 if (po->btj != NULL) {
8726 // jumptable
8727 for (j = 0; j < po->btj->count; j++) {
8728 check_i(po, po->btj->d[j].bt_i);
8729 gen_hdr_dep_pass(po->btj->d[j].bt_i, opcnt, cbits, fp,
8730 regmask_save, regmask_dst, regmask_dep, regmask_use,
8731 has_ret);
8732 }
8733 return;
8734 }
8735
8736 check_i(po, po->bt_i);
8737 if (po->flags & OPF_CJMP) {
8738 gen_hdr_dep_pass(po->bt_i, opcnt, cbits, fp,
8739 regmask_save, regmask_dst, regmask_dep, regmask_use,
8740 has_ret);
8741 }
8742 else {
8743 i = po->bt_i - 1;
8744 }
8745 continue;
8746 }
8747
8748 if (po->flags & OPF_FARG)
8749 /* (just calculate register deps) */;
8750 else if (po->op == OP_PUSH && po->operand[0].type == OPT_REG)
8751 {
8752 reg = po->operand[0].reg;
8753 ferr_assert(po, reg >= 0);
8754
8755 if (po->flags & OPF_RSAVE) {
8756 regmask_save |= 1 << reg;
8757 continue;
8758 }
8759 if (po->flags & OPF_DONE)
8760 continue;
8761
8762 ret = scan_for_pop(i + 1, opcnt, i + opcnt * 2,
8763 reg, 0, 0, 0, 0);
8764 if (ret == 1) {
8765 regmask_save |= 1 << reg;
8766 po->flags |= OPF_RMD;
8767 scan_for_pop(i + 1, opcnt, i + opcnt * 3,
8768 reg, 0, 0, 0, OPF_RMD);
8769 continue;
8770 }
8771 }
8772 else if (po->flags & OPF_RMD)
8773 continue;
8774 else if (po->op == OP_CALL) {
8775 po->regmask_dst |= 1 << xAX;
8776
8777 dep = hg_fp_find_dep(fp, po->operand[0].name);
8778 if (dep != NULL) {
8779 dep->regmask_live = regmask_save | regmask_dst;
8780 if (g_bp_frame && !(po->flags & OPF_EBP_S))
8781 dep->regmask_live |= 1 << xBP;
8782 }
8783 if ((po->flags & OPF_TAIL) && po->pp != NULL
8784 && po->pp->is_stdcall)
8785 fp->is_stdcall = 1;
8786 }
8787 else if (po->op == OP_RET) {
8788 if (po->operand_cnt > 0) {
8789 fp->is_stdcall = 1;
8790 if (fp->argc_stack >= 0
8791 && fp->argc_stack != po->operand[0].val / 4)
8792 ferr(po, "ret mismatch? (%d)\n", fp->argc_stack * 4);
8793 fp->argc_stack = po->operand[0].val / 4;
8794 }
8795 }
8796
8797 if (!fp->eax_pass && (po->flags & OPF_TAIL)) {
8798 if (po->op == OP_CALL) {
8799 j = i;
8800 ret = 1;
8801 }
8802 else {
8803 j = -1;
8804 from_caller = 0;
8805 ret = resolve_origin_reg(i, xAX, i + opcnt * 4, &j, &from_caller);
8806 }
8807
8808 if (ret != 1 && from_caller) {
8809 // unresolved eax - probably void func
8810 *has_ret = 0;
8811 fp->eax_pass = 1;
8812 }
8813 else {
8814 if (j >= 0 && ops[j].op == OP_CALL) {
8815 if (ops[j].pp != NULL && !ops[j].pp->is_unresolved) {
8816 int call_has_ret = !IS(ops[j].pp->ret_type.name, "void");
8817 if (ops[j].pp->is_noreturn) {
8818 // could be some fail path
8819 if (*has_ret == -1)
8820 *has_ret = call_has_ret;
8821 }
8822 else
8823 *has_ret = call_has_ret;
8824 }
8825 else {
8826 dep = hg_fp_find_dep(fp, ops[j].operand[0].name);
8827 if (dep != NULL)
8828 dep->ret_dep = 1;
8829 else
8830 *has_ret = 1;
8831 }
8832 }
8833 else
8834 *has_ret = 1;
8835 }
8836 }
8837
8838 l = regmask_save | regmask_dst;
8839 if (g_bp_frame && !(po->flags & OPF_EBP_S))
8840 l |= 1 << xBP;
8841
8842 l = po->regmask_src & ~l;
8843#if 0
8844 if (l)
8845 fnote(po, "dep |= %04x, dst %04x, save %04x (f %x)\n",
8846 l, regmask_dst, regmask_save, po->flags);
8847#endif
8848 *regmask_dep |= l;
8849 *regmask_use |= (po->regmask_src | po->regmask_dst)
8850 & ~regmask_save;
8851 regmask_dst |= po->regmask_dst;
8852
8853 if (po->flags & OPF_TAIL) {
8854 if (!(po->flags & OPF_CC)) // not cond. tailcall
8855 return;
8856 }
8857 }
8858}
8859
8860static void gen_hdr(const char *funcn, int opcnt)
8861{
8862 unsigned char cbits[MAX_OPS / 8];
8863 const struct parsed_proto *pp_c;
8864 struct parsed_proto *pp;
8865 struct func_prototype *fp;
8866 struct func_proto_dep *dep;
8867 struct parsed_op *po;
8868 const char *tmpname;
8869 int regmask_dummy = 0;
8870 int regmask_dep;
8871 int regmask_use;
8872 int max_bp_offset = 0;
8873 int has_ret;
8874 int i, j, l;
8875 int ret;
8876
8877 pp_c = proto_parse(g_fhdr, funcn, 1);
8878 if (pp_c != NULL)
8879 // already in seed, will add to hg_fp later
8880 return;
8881
8882 fp = hg_fp_add(funcn);
8883
8884 g_bp_frame = g_sp_frame = g_stack_fsz = 0;
8885 g_stack_frame_used = 0;
8886 g_seh_size = 0;
8887
8888 // pass1:
8889 // - resolve all branches
8890 // - parse calls with labels
8891 resolve_branches_parse_calls(opcnt);
8892
8893 // pass2:
8894 // - handle ebp/esp frame, remove ops related to it
8895 scan_prologue_epilogue(opcnt, NULL);
8896
8897 // pass3:
8898 // - remove dead labels
8899 // - collect calls
8900 // - collect function ptr refs
8901 for (i = 0; i < opcnt; i++)
8902 {
8903 if (g_labels[i] != NULL && g_label_refs[i].i == -1) {
8904 free(g_labels[i]);
8905 g_labels[i] = NULL;
8906 }
8907
8908 po = &ops[i];
8909 if (po->flags & (OPF_RMD|OPF_DONE))
8910 continue;
8911
8912 if (po->op == OP_CALL) {
8913 if (po->operand[0].type == OPT_LABEL)
8914 hg_fp_add_dep(fp, opr_name(po, 0), 0);
8915 else if (po->pp != NULL)
8916 hg_fp_add_dep(fp, po->pp->name, 0);
8917 }
8918 else if (po->op == OP_MOV && po->operand[1].type == OPT_OFFSET) {
8919 tmpname = opr_name(po, 1);
8920 if (IS_START(tmpname, "p_") || IS_START(tmpname, "sub_"))
8921 hg_fp_add_dep(fp, tmpname, 1);
8922 }
8923 else if (po->op == OP_PUSH && po->operand[0].type == OPT_OFFSET) {
8924 tmpname = opr_name(po, 0);
8925 if (IS_START(tmpname, "p_") || IS_START(tmpname, "sub_"))
8926 hg_fp_add_dep(fp, tmpname, 1);
8927 }
8928 }
8929
8930 // pass4:
8931 // - handle push <const>/pop pairs
8932 for (i = 0; i < opcnt; i++)
8933 {
8934 po = &ops[i];
8935 if (po->flags & (OPF_RMD|OPF_DONE))
8936 continue;
8937
8938 if (po->op == OP_PUSH && po->operand[0].type == OPT_CONST)
8939 scan_for_pop_const(i, opcnt, i + opcnt * 13);
8940 }
8941
8942 // pass5:
8943 // - process trivial calls
8944 for (i = 0; i < opcnt; i++)
8945 {
8946 po = &ops[i];
8947 if (po->flags & (OPF_RMD|OPF_DONE))
8948 continue;
8949
8950 if (po->op == OP_CALL)
8951 {
8952 pp = process_call_early(i, opcnt, &j);
8953 if (pp != NULL) {
8954 if (!(po->flags & OPF_ATAIL))
8955 // since we know the args, try to collect them
8956 if (collect_call_args_early(i, opcnt, pp, NULL, NULL) != 0)
8957 pp = NULL;
8958 }
8959
8960 if (pp != NULL) {
8961 if (j >= 0) {
8962 // commit esp adjust
8963 if (ops[j].op != OP_POP)
8964 patch_esp_adjust(&ops[j], pp->argc_stack * 4);
8965 else {
8966 for (l = 0; l < pp->argc_stack; l++)
8967 ops[j + l].flags |= OPF_DONE | OPF_RMD | OPF_NOREGS;
8968 }
8969 }
8970
8971 po->flags |= OPF_DONE;
8972 }
8973 }
8974 }
8975
8976 // pass6:
8977 // - track saved regs (simple)
8978 // - process calls
8979 for (i = 0; i < opcnt; i++)
8980 {
8981 po = &ops[i];
8982 if (po->flags & (OPF_RMD|OPF_DONE))
8983 continue;
8984
8985 if (po->op == OP_PUSH && po->operand[0].type == OPT_REG
8986 && po->operand[0].reg != xCX)
8987 {
8988 ret = scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg, 0);
8989 if (ret == 1) {
8990 // regmask_save |= 1 << po->operand[0].reg; // do it later
8991 po->flags |= OPF_RSAVE | OPF_RMD | OPF_DONE;
8992 scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg, OPF_RMD);
8993 }
8994 }
8995 else if (po->op == OP_CALL)
8996 {
8997 pp = process_call(i, opcnt);
8998
8999 if (!pp->is_unresolved && !(po->flags & OPF_ATAIL)) {
9000 // since we know the args, collect them
9001 ret = collect_call_args(po, i, opcnt, pp, &regmask_dummy,
9002 i + opcnt * 1);
9003 }
9004 if (!(po->flags & OPF_TAIL)
9005 && po->operand[0].type == OPT_LABEL)
9006 {
9007 dep = hg_fp_find_dep(fp, opr_name(po, 0));
9008 ferr_assert(po, dep != NULL);
9009 // treat al write as overwrite to avoid many false positives
9010 find_next_read_reg(i + 1, opcnt, xAX, OPLM_BYTE,
9011 i + opcnt * 25, &j);
9012 if (j != -1)
9013 dep->has_ret = 1;
9014 find_next_read_reg(i + 1, opcnt, xDX, OPLM_BYTE,
9015 i + opcnt * 26, &j);
9016 if (j != -1 && !IS_OP_INDIRECT_CALL(&ops[j]))
9017 dep->has_ret64 = 1;
9018 }
9019 }
9020 }
9021
9022 // pass7
9023 memset(cbits, 0, (opcnt + 7) / 8);
9024 regmask_dep = regmask_use = 0;
9025 has_ret = -1;
9026
9027 gen_hdr_dep_pass(0, opcnt, cbits, fp, 0, 0,
9028 &regmask_dep, &regmask_use, &has_ret);
9029
9030 // find unreachable code - must be fixed in IDA
9031 for (i = 0; i < opcnt; i++)
9032 {
9033 if (cbits[i >> 3] & (1 << (i & 7)))
9034 continue;
9035
9036 if (g_labels[i] == NULL && i > 0 && ops[i - 1].op == OP_CALL
9037 && ops[i - 1].pp != NULL && ops[i - 1].pp->is_osinc)
9038 {
9039 // the compiler sometimes still generates code after
9040 // noreturn OS functions
9041 break;
9042 }
9043 if (!(ops[i].flags & OPF_RMD)
9044 && ops[i].op != OP_NOP && ops[i].op != OPP_ABORT)
9045 {
9046 ferr(&ops[i], "unreachable code\n");
9047 }
9048 }
9049
9050 for (i = 0; i < g_eqcnt; i++) {
9051 if (g_eqs[i].offset > max_bp_offset && g_eqs[i].offset < 4*32)
9052 max_bp_offset = g_eqs[i].offset;
9053 }
9054
9055 if (fp->argc_stack < 0) {
9056 max_bp_offset = (max_bp_offset + 3) & ~3;
9057 fp->argc_stack = max_bp_offset / 4;
9058 if ((g_ida_func_attr & IDAFA_BP_FRAME) && fp->argc_stack > 0)
9059 fp->argc_stack--;
9060 }
9061
9062 fp->regmask_dep = regmask_dep & ~((1 << xSP) | mxSTa);
9063 fp->regmask_use = regmask_use;
9064 fp->has_ret = has_ret;
9065#if 0
9066 printf("// has_ret %d, regmask_dep %x\n",
9067 fp->has_ret, fp->regmask_dep);
9068 output_hdr_fp(stdout, fp, 1);
9069 if (IS(funcn, "sub_10007F72")) exit(1);
9070#endif
9071
9072 gen_x_cleanup(opcnt);
9073}
9074
9075static void hg_fp_resolve_deps(struct func_prototype *fp)
9076{
9077 struct func_prototype fp_s;
9078 struct func_proto_dep *dep;
9079 int regmask_dep;
9080 int i;
9081
9082 // this thing is recursive, so mark first..
9083 fp->dep_resolved = 1;
9084
9085 for (i = 0; i < fp->dep_func_cnt; i++) {
9086 dep = &fp->dep_func[i];
9087
9088 strcpy(fp_s.name, dep->name);
9089 dep->proto = bsearch(&fp_s, hg_fp, hg_fp_cnt,
9090 sizeof(hg_fp[0]), hg_fp_cmp_name);
9091 if (dep->proto != NULL) {
9092 if (dep->ptr_taken) {
9093 dep->proto->ptr_taken = 1;
9094 continue;
9095 }
9096
9097 if (!dep->proto->dep_resolved)
9098 hg_fp_resolve_deps(dep->proto);
9099
9100 regmask_dep = ~dep->regmask_live
9101 & dep->proto->regmask_dep;
9102 fp->regmask_dep |= regmask_dep;
9103 // printf("dep %s %s |= %x\n", fp->name,
9104 // fp->dep_func[i].name, regmask_dep);
9105
9106 if (dep->has_ret && (dep->proto->regmask_use & mxAX))
9107 dep->proto->has_ret = 1;
9108 if (dep->has_ret64 && (dep->proto->regmask_use & mxDX))
9109 dep->proto->has_ret64 = 1;
9110 if (fp->has_ret == -1 && dep->ret_dep)
9111 fp->has_ret = dep->proto->has_ret;
9112 }
9113 }
9114}
9115
9116// make all thiscall/edx arg functions referenced from .data fastcall
9117static void do_func_refs_from_data(void)
9118{
9119 struct func_prototype *fp, fp_s;
9120 int i;
9121
9122 for (i = 0; i < hg_ref_cnt; i++) {
9123 strcpy(fp_s.name, hg_refs[i]);
9124 fp = bsearch(&fp_s, hg_fp, hg_fp_cnt,
9125 sizeof(hg_fp[0]), hg_fp_cmp_name);
9126 if (fp != NULL)
9127 fp->ptr_taken = 1;
9128 }
9129}
9130
9131static void output_hdr_fp(FILE *fout, const struct func_prototype *fp,
9132 int count)
9133{
9134 const struct parsed_proto *pp;
9135 char *p, namebuf[NAMELEN];
9136 const char *name;
9137 int regmask_dep;
9138 int argc_normal;
9139 int j, arg;
9140
9141 for (; count > 0; count--, fp++) {
9142 if (fp->has_ret == -1)
9143 fprintf(fout, "// ret unresolved\n");
9144#if 0
9145 fprintf(fout, "// dep:");
9146 for (j = 0; j < fp->dep_func_cnt; j++) {
9147 fprintf(fout, " %s/", fp->dep_func[j].name);
9148 if (fp->dep_func[j].proto != NULL)
9149 fprintf(fout, "%04x/%d", fp->dep_func[j].proto->regmask_dep,
9150 fp->dep_func[j].proto->has_ret);
9151 }
9152 fprintf(fout, "\n");
9153#endif
9154
9155 p = strchr(fp->name, '@');
9156 if (p != NULL) {
9157 memcpy(namebuf, fp->name, p - fp->name);
9158 namebuf[p - fp->name] = 0;
9159 name = namebuf;
9160 }
9161 else
9162 name = fp->name;
9163 if (name[0] == '_')
9164 name++;
9165
9166 pp = proto_parse(g_fhdr, name, 1);
9167 if (pp != NULL && pp->is_include)
9168 continue;
9169
9170 if (fp->pp != NULL) {
9171 // part of seed, output later
9172 continue;
9173 }
9174
9175 regmask_dep = fp->regmask_dep;
9176 argc_normal = fp->argc_stack;
9177 if (fp->ptr_taken && regmask_dep
9178 && (regmask_dep & ~(mxCX|mxDX)) == 0)
9179 {
9180 if ((regmask_dep & mxDX) || fp->argc_stack > 0)
9181 regmask_dep |= mxCX | mxDX;
9182 }
9183
9184 fprintf(fout, "%-5s",
9185 fp->pp ? fp->pp->ret_type.name :
9186 fp->has_ret64 ? "__int64" :
9187 fp->has_ret ? "int" : "void");
9188 if (regmask_dep == mxCX && fp->is_stdcall && fp->argc_stack > 0) {
9189 fprintf(fout, "/*__thiscall*/ ");
9190 argc_normal++;
9191 regmask_dep = 0;
9192 }
9193 else if ((regmask_dep == (mxCX | mxDX)
9194 && (fp->is_stdcall || fp->argc_stack == 0))
9195 || (regmask_dep == mxCX && fp->argc_stack == 0))
9196 {
9197 fprintf(fout, " __fastcall ");
9198 if (!(regmask_dep & (1 << xDX)) && fp->argc_stack == 0)
9199 argc_normal = 1;
9200 else
9201 argc_normal += 2;
9202 regmask_dep = 0;
9203 }
9204 else if (regmask_dep && !fp->is_stdcall) {
9205 fprintf(fout, "/*__usercall*/ ");
9206 }
9207 else if (regmask_dep) {
9208 fprintf(fout, "/*__userpurge*/ ");
9209 }
9210 else if (fp->is_stdcall)
9211 fprintf(fout, " __stdcall ");
9212 else
9213 fprintf(fout, " __cdecl ");
9214
9215 fprintf(fout, "%s(", name);
9216
9217 arg = 0;
9218 for (j = 0; j < xSP; j++) {
9219 if (regmask_dep & (1 << j)) {
9220 arg++;
9221 if (arg != 1)
9222 fprintf(fout, ", ");
9223 if (fp->pp != NULL)
9224 fprintf(fout, "%s", fp->pp->arg[arg - 1].type.name);
9225 else
9226 fprintf(fout, "int");
9227 fprintf(fout, " a%d/*<%s>*/", arg, regs_r32[j]);
9228 }
9229 }
9230
9231 for (j = 0; j < argc_normal; j++) {
9232 arg++;
9233 if (arg != 1)
9234 fprintf(fout, ", ");
9235 if (fp->pp != NULL) {
9236 fprintf(fout, "%s", fp->pp->arg[arg - 1].type.name);
9237 if (!fp->pp->arg[arg - 1].type.is_ptr)
9238 fprintf(fout, " ");
9239 }
9240 else
9241 fprintf(fout, "int ");
9242 fprintf(fout, "a%d", arg);
9243 }
9244
9245 fprintf(fout, ");\n");
9246 }
9247}
9248
9249static void output_hdr(FILE *fout)
9250{
9251 static const char *lmod_c_names[] = {
9252 [OPLM_UNSPEC] = "???",
9253 [OPLM_BYTE] = "uint8_t",
9254 [OPLM_WORD] = "uint16_t",
9255 [OPLM_DWORD] = "uint32_t",
9256 [OPLM_QWORD] = "uint64_t",
9257 };
9258 const struct scanned_var *var;
9259 struct func_prototype *fp;
9260 char line[256] = { 0, };
9261 char name[256];
9262 int i;
9263
9264 // add stuff from headers
9265 for (i = 0; i < pp_cache_size; i++) {
9266 if (pp_cache[i].is_cinc && !pp_cache[i].is_stdcall)
9267 snprintf(name, sizeof(name), "_%s", pp_cache[i].name);
9268 else
9269 snprintf(name, sizeof(name), "%s", pp_cache[i].name);
9270 fp = hg_fp_add(name);
9271 fp->pp = &pp_cache[i];
9272 fp->argc_stack = fp->pp->argc_stack;
9273 fp->is_stdcall = fp->pp->is_stdcall;
9274 fp->regmask_dep = get_pp_arg_regmask_src(fp->pp);
9275 fp->has_ret = !IS(fp->pp->ret_type.name, "void");
9276 }
9277
9278 // resolve deps
9279 qsort(hg_fp, hg_fp_cnt, sizeof(hg_fp[0]), hg_fp_cmp_name);
9280 for (i = 0; i < hg_fp_cnt; i++)
9281 hg_fp_resolve_deps(&hg_fp[i]);
9282
9283 // adjust functions referenced from data segment
9284 do_func_refs_from_data();
9285
9286 // final adjustments
9287 for (i = 0; i < hg_fp_cnt; i++) {
9288 if (hg_fp[i].eax_pass && (hg_fp[i].regmask_dep & mxAX))
9289 hg_fp[i].has_ret = 1;
9290 }
9291
9292 // note: messes up .proto ptr, don't use
9293 //qsort(hg_fp, hg_fp_cnt, sizeof(hg_fp[0]), hg_fp_cmp_id);
9294
9295 // output variables
9296 for (i = 0; i < hg_var_cnt; i++) {
9297 var = &hg_vars[i];
9298
9299 if (var->pp != NULL)
9300 // part of seed
9301 continue;
9302 else if (var->is_c_str)
9303 fprintf(fout, "extern %-8s %s[];", "char", var->name);
9304 else
9305 fprintf(fout, "extern %-8s %s;",
9306 lmod_c_names[var->lmod], var->name);
9307
9308 if (var->is_seeded)
9309 fprintf(fout, " // seeded");
9310 fprintf(fout, "\n");
9311 }
9312
9313 fprintf(fout, "\n");
9314
9315 // output function prototypes
9316 output_hdr_fp(fout, hg_fp, hg_fp_cnt);
9317
9318 // seed passthrough
9319 fprintf(fout, "\n// - seed -\n");
9320
9321 rewind(g_fhdr);
9322 while (fgets(line, sizeof(line), g_fhdr))
9323 fwrite(line, 1, strlen(line), fout);
9324}
9325
9326// '=' needs special treatment
9327// also ' quote
9328static char *next_word_s(char *w, size_t wsize, char *s)
9329{
9330 size_t i;
9331
9332 s = sskip(s);
9333
9334 i = 0;
9335 if (*s == '\'' && s[1] != '\r' && s[1] != '\n') {
9336 w[0] = s[0];
9337 for (i = 1; i < wsize - 1; i++) {
9338 if (s[i] == 0) {
9339 printf("warning: missing closing quote: \"%s\"\n", s);
9340 break;
9341 }
9342 if (s[i] == '\'')
9343 break;
9344 w[i] = s[i];
9345 }
9346 }
9347
9348 for (; i < wsize - 1; i++) {
9349 if (s[i] == 0 || my_isblank(s[i]) || (s[i] == '=' && i > 0))
9350 break;
9351 w[i] = s[i];
9352 }
9353 w[i] = 0;
9354
9355 if (s[i] != 0 && !my_isblank(s[i]) && s[i] != '=')
9356 printf("warning: '%s' truncated\n", w);
9357
9358 return s + i;
9359}
9360
9361static int cmpstringp(const void *p1, const void *p2)
9362{
9363 return strcmp(*(char * const *)p1, *(char * const *)p2);
9364}
9365
9366static int is_xref_needed(char *p, char **rlist, int rlist_len)
9367{
9368 char *p2;
9369
9370 p = sskip(p);
9371 if (strstr(p, "..."))
9372 // unable to determine, assume needed
9373 return 1;
9374
9375 if (*p == '.') // .text, .data, ...
9376 // ref from other data or non-function -> no
9377 return 0;
9378
9379 p2 = strpbrk(p, "+:\r\n\x18");
9380 if (p2 != NULL)
9381 *p2 = 0;
9382 if (bsearch(&p, rlist, rlist_len, sizeof(rlist[0]), cmpstringp))
9383 // referenced from removed code
9384 return 0;
9385
9386 return 1;
9387}
9388
9389static int ida_xrefs_show_need(FILE *fasm, char *p,
9390 char **rlist, int rlist_len)
9391{
9392 int found_need = 0;
9393 char line[256];
9394 long pos;
9395
9396 p = strrchr(p, ';');
9397 if (p != NULL && *p == ';') {
9398 if (IS_START(p + 2, "sctref"))
9399 return 1;
9400 if (IS_START(p + 2, "DATA XREF: ")) {
9401 p += 13;
9402 if (is_xref_needed(p, rlist, rlist_len))
9403 return 1;
9404 }
9405 }
9406
9407 pos = ftell(fasm);
9408 while (1)
9409 {
9410 if (!my_fgets(line, sizeof(line), fasm))
9411 break;
9412 // non-first line is always indented
9413 if (!my_isblank(line[0]))
9414 break;
9415
9416 // should be no content, just comment
9417 p = sskip(line);
9418 if (*p != ';')
9419 break;
9420
9421 p = strrchr(p, ';');
9422 p += 2;
9423
9424 if (IS_START(p, "sctref")) {
9425 found_need = 1;
9426 break;
9427 }
9428
9429 // it's printed once, but no harm to check again
9430 if (IS_START(p, "DATA XREF: "))
9431 p += 11;
9432
9433 if (is_xref_needed(p, rlist, rlist_len)) {
9434 found_need = 1;
9435 break;
9436 }
9437 }
9438 fseek(fasm, pos, SEEK_SET);
9439 return found_need;
9440}
9441
9442static void scan_variables(FILE *fasm, char **rlist, int rlist_len)
9443{
9444 struct scanned_var *var;
9445 char line[256] = { 0, };
9446 char words[4][256];
9447 int no_identifier;
9448 char *p = NULL;
9449 int wordc;
9450 int l;
9451
9452 while (!feof(fasm))
9453 {
9454 // skip to next data section
9455 while (my_fgets(line, sizeof(line), fasm))
9456 {
9457 asmln++;
9458
9459 p = sskip(line);
9460 if (*p == 0 || *p == ';')
9461 continue;
9462
9463 p = sskip(next_word_s(words[0], sizeof(words[0]), p));
9464 if (*p == 0 || *p == ';')
9465 continue;
9466
9467 if (*p != 's' || !IS_START(p, "segment para public"))
9468 continue;
9469
9470 break;
9471 }
9472
9473 if (p == NULL || !IS_START(p, "segment para public"))
9474 break;
9475 p = sskip(p + 19);
9476
9477 if (!IS_START(p, "'DATA'"))
9478 continue;
9479
9480 // now process it
9481 while (my_fgets(line, sizeof(line), fasm))
9482 {
9483 asmln++;
9484
9485 p = line;
9486 no_identifier = my_isblank(*p);
9487
9488 p = sskip(p);
9489 if (*p == 0 || *p == ';')
9490 continue;
9491
9492 for (wordc = 0; wordc < ARRAY_SIZE(words); wordc++) {
9493 words[wordc][0] = 0;
9494 p = sskip(next_word_s(words[wordc], sizeof(words[0]), p));
9495 if (*p == 0 || *p == ';') {
9496 wordc++;
9497 break;
9498 }
9499 }
9500
9501 if (wordc == 2 && IS(words[1], "ends"))
9502 break;
9503 if (wordc < 2)
9504 continue;
9505
9506 if (no_identifier) {
9507 if (wordc >= 3 && IS(words[0], "dd") && IS(words[1], "offset"))
9508 hg_ref_add(words[2]);
9509 continue;
9510 }
9511
9512 if (IS_START(words[0], "__IMPORT_DESCRIPTOR_")) {
9513 // when this starts, we don't need anything from this section
9514 break;
9515 }
9516
9517 // check refs comment(s)
9518 if (!ida_xrefs_show_need(fasm, p, rlist, rlist_len))
9519 continue;
9520
9521 if ((hg_var_cnt & 0xff) == 0) {
9522 hg_vars = realloc(hg_vars, sizeof(hg_vars[0])
9523 * (hg_var_cnt + 0x100));
9524 my_assert_not(hg_vars, NULL);
9525 memset(hg_vars + hg_var_cnt, 0, sizeof(hg_vars[0]) * 0x100);
9526 }
9527
9528 var = &hg_vars[hg_var_cnt++];
9529 snprintf(var->name, sizeof(var->name), "%s", words[0]);
9530
9531 // maybe already in seed header?
9532 var->pp = proto_parse(g_fhdr, var->name, 1);
9533 if (var->pp != NULL) {
9534 if (var->pp->is_fptr) {
9535 var->lmod = OPLM_DWORD;
9536 //var->is_ptr = 1;
9537 }
9538 else if (var->pp->is_func)
9539 aerr("func?\n");
9540 else if (!guess_lmod_from_c_type(&var->lmod, &var->pp->type))
9541 aerr("unhandled C type '%s' for '%s'\n",
9542 var->pp->type.name, var->name);
9543
9544 var->is_seeded = 1;
9545 continue;
9546 }
9547
9548 if (IS(words[1], "dd")) {
9549 var->lmod = OPLM_DWORD;
9550 if (wordc >= 4 && IS(words[2], "offset"))
9551 hg_ref_add(words[3]);
9552 }
9553 else if (IS(words[1], "dw"))
9554 var->lmod = OPLM_WORD;
9555 else if (IS(words[1], "db")) {
9556 var->lmod = OPLM_BYTE;
9557 if (wordc >= 3 && (l = strlen(words[2])) > 4) {
9558 if (words[2][0] == '\'' && IS(words[2] + l - 2, ",0"))
9559 var->is_c_str = 1;
9560 }
9561 }
9562 else if (IS(words[1], "dq"))
9563 var->lmod = OPLM_QWORD;
9564 //else if (IS(words[1], "dt"))
9565 else
9566 aerr("type '%s' not known\n", words[1]);
9567 }
9568 }
9569
9570 rewind(fasm);
9571 asmln = 0;
9572}
9573
9574static void set_label(int i, const char *name)
9575{
9576 const char *p;
9577 int len;
9578
9579 len = strlen(name);
9580 p = strchr(name, ':');
9581 if (p != NULL)
9582 len = p - name;
9583
9584 if (g_labels[i] != NULL && !IS_START(g_labels[i], "algn_"))
9585 aerr("dupe label '%s' vs '%s'?\n", name, g_labels[i]);
9586 g_labels[i] = realloc(g_labels[i], len + 1);
9587 my_assert_not(g_labels[i], NULL);
9588 memcpy(g_labels[i], name, len);
9589 g_labels[i][len] = 0;
9590}
9591
9592struct chunk_item {
9593 char *name;
9594 long fptr;
9595 int asmln;
9596};
9597
9598static struct chunk_item *func_chunks;
9599static int func_chunk_cnt;
9600static int func_chunk_alloc;
9601
9602static void add_func_chunk(FILE *fasm, const char *name, int line)
9603{
9604 if (func_chunk_cnt >= func_chunk_alloc) {
9605 func_chunk_alloc *= 2;
9606 func_chunks = realloc(func_chunks,
9607 func_chunk_alloc * sizeof(func_chunks[0]));
9608 my_assert_not(func_chunks, NULL);
9609 }
9610 func_chunks[func_chunk_cnt].fptr = ftell(fasm);
9611 func_chunks[func_chunk_cnt].name = strdup(name);
9612 func_chunks[func_chunk_cnt].asmln = line;
9613 func_chunk_cnt++;
9614}
9615
9616static int cmp_chunks(const void *p1, const void *p2)
9617{
9618 const struct chunk_item *c1 = p1, *c2 = p2;
9619 return strcmp(c1->name, c2->name);
9620}
9621
9622static void scan_ahead_for_chunks(FILE *fasm)
9623{
9624 char words[2][256];
9625 char line[256];
9626 long oldpos;
9627 int oldasmln;
9628 int wordc;
9629 char *p;
9630 int i;
9631
9632 oldpos = ftell(fasm);
9633 oldasmln = asmln;
9634
9635 while (my_fgets(line, sizeof(line), fasm))
9636 {
9637 wordc = 0;
9638 asmln++;
9639
9640 p = sskip(line);
9641 if (*p == 0)
9642 continue;
9643
9644 if (*p == ';')
9645 {
9646 // get rid of random tabs
9647 for (i = 0; line[i] != 0; i++)
9648 if (line[i] == '\t')
9649 line[i] = ' ';
9650
9651 if (p[2] == 'S' && IS_START(p, "; START OF FUNCTION CHUNK FOR "))
9652 {
9653 p += 30;
9654 next_word(words[0], sizeof(words[0]), p);
9655 if (words[0][0] == 0)
9656 aerr("missing name for func chunk?\n");
9657
9658 add_func_chunk(fasm, words[0], asmln);
9659 }
9660 else if (IS_START(p, "; sctend"))
9661 break;
9662
9663 continue;
9664 } // *p == ';'
9665
9666 for (wordc = 0; wordc < ARRAY_SIZE(words); wordc++) {
9667 words[wordc][0] = 0;
9668 p = sskip(next_word_s(words[wordc], sizeof(words[0]), p));
9669 if (*p == 0 || *p == ';') {
9670 wordc++;
9671 break;
9672 }
9673 }
9674
9675 if (wordc == 2 && IS(words[1], "ends"))
9676 break;
9677 }
9678
9679 fseek(fasm, oldpos, SEEK_SET);
9680 asmln = oldasmln;
9681}
9682
9683int main(int argc, char *argv[])
9684{
9685 FILE *fout, *fasm, *frlist;
9686 struct parsed_data *pd = NULL;
9687 int pd_alloc = 0;
9688 char **rlist = NULL;
9689 int rlist_len = 0;
9690 int rlist_alloc = 0;
9691 int func_chunks_used = 0;
9692 int func_chunks_sorted = 0;
9693 int func_chunk_i = -1;
9694 long func_chunk_ret = 0;
9695 int func_chunk_ret_ln = 0;
9696 int scanned_ahead = 0;
9697 char line[256];
9698 char words[20][256];
9699 enum opr_lenmod lmod;
9700 char *sctproto = NULL;
9701 int in_func = 0;
9702 int pending_endp = 0;
9703 int skip_code = 0;
9704 int skip_code_end = 0;
9705 int skip_warned = 0;
9706 int eq_alloc;
9707 int verbose = 0;
9708 int multi_seg = 0;
9709 int end = 0;
9710 int arg_out;
9711 int arg;
9712 int pi = 0;
9713 int i, j;
9714 int ret, len;
9715 char *p, *p2;
9716 int wordc;
9717
9718 for (arg = 1; arg < argc; arg++) {
9719 if (IS(argv[arg], "-v"))
9720 verbose = 1;
9721 else if (IS(argv[arg], "-rf"))
9722 g_allow_regfunc = 1;
9723 else if (IS(argv[arg], "-uc"))
9724 g_allow_user_icall = 1;
9725 else if (IS(argv[arg], "-wu"))
9726 g_nowarn_reguse = 1;
9727 else if (IS(argv[arg], "-m"))
9728 multi_seg = 1;
9729 else if (IS(argv[arg], "-hdr"))
9730 g_header_mode = g_quiet_pp = g_allow_regfunc = 1;
9731 else
9732 break;
9733 }
9734
9735 if (argc < arg + 3) {
9736 printf("usage:\n%s [options] <.c> <.asm> <hdr.h> [rlist]*\n"
9737 "%s -hdr <out.h> <.asm> <seed.h> [rlist]*\n"
9738 "options:\n"
9739 " -hdr - header generation mode\n"
9740 " -rf - allow unannotated indirect calls\n"
9741 " -uc - allow ind. calls/refs to __usercall\n"
9742 " -m - allow multiple .text sections\n"
9743 " -wu - don't warn about bad reg use\n"
9744 "[rlist] is a file with function names to skip,"
9745 " one per line\n",
9746 argv[0], argv[0]);
9747 return 1;
9748 }
9749
9750 arg_out = arg++;
9751
9752 asmfn = argv[arg++];
9753 fasm = fopen(asmfn, "r");
9754 my_assert_not(fasm, NULL);
9755
9756 hdrfn = argv[arg++];
9757 g_fhdr = fopen(hdrfn, "r");
9758 my_assert_not(g_fhdr, NULL);
9759
9760 rlist_alloc = 64;
9761 rlist = malloc(rlist_alloc * sizeof(rlist[0]));
9762 my_assert_not(rlist, NULL);
9763 // needs special handling..
9764 rlist[rlist_len++] = "__alloca_probe";
9765
9766 func_chunk_alloc = 32;
9767 func_chunks = malloc(func_chunk_alloc * sizeof(func_chunks[0]));
9768 my_assert_not(func_chunks, NULL);
9769
9770 memset(words, 0, sizeof(words));
9771
9772 for (; arg < argc; arg++) {
9773 int skip_func = 0;
9774
9775 frlist = fopen(argv[arg], "r");
9776 my_assert_not(frlist, NULL);
9777
9778 while (my_fgets(line, sizeof(line), frlist)) {
9779 p = sskip(line);
9780 if (*p == 0 || *p == ';')
9781 continue;
9782 if (*p == '#') {
9783 if (IS_START(p, "#if 0")
9784 || (g_allow_regfunc && IS_START(p, "#if NO_REGFUNC")))
9785 {
9786 skip_func = 1;
9787 }
9788 else if (IS_START(p, "#endif"))
9789 skip_func = 0;
9790 continue;
9791 }
9792 if (skip_func)
9793 continue;
9794
9795 p = next_word(words[0], sizeof(words[0]), p);
9796 if (words[0][0] == 0)
9797 continue;
9798
9799 if (rlist_len >= rlist_alloc) {
9800 rlist_alloc = rlist_alloc * 2 + 64;
9801 rlist = realloc(rlist, rlist_alloc * sizeof(rlist[0]));
9802 my_assert_not(rlist, NULL);
9803 }
9804 rlist[rlist_len++] = strdup(words[0]);
9805 }
9806
9807 fclose(frlist);
9808 frlist = NULL;
9809 }
9810
9811 if (rlist_len > 0)
9812 qsort(rlist, rlist_len, sizeof(rlist[0]), cmpstringp);
9813
9814 fout = fopen(argv[arg_out], "w");
9815 my_assert_not(fout, NULL);
9816
9817 eq_alloc = 128;
9818 g_eqs = malloc(eq_alloc * sizeof(g_eqs[0]));
9819 my_assert_not(g_eqs, NULL);
9820
9821 for (i = 0; i < ARRAY_SIZE(g_label_refs); i++) {
9822 g_label_refs[i].i = -1;
9823 g_label_refs[i].next = NULL;
9824 }
9825
9826 if (g_header_mode)
9827 scan_variables(fasm, rlist, rlist_len);
9828
9829 while (my_fgets(line, sizeof(line), fasm))
9830 {
9831 wordc = 0;
9832 asmln++;
9833
9834 p = sskip(line);
9835 if (*p == 0)
9836 continue;
9837
9838 // get rid of random tabs
9839 for (i = 0; line[i] != 0; i++)
9840 if (line[i] == '\t')
9841 line[i] = ' ';
9842
9843 if (*p == ';')
9844 {
9845 if (p[2] == '=' && IS_START(p, "; =============== S U B"))
9846 goto do_pending_endp; // eww..
9847
9848 if (p[2] == 'A' && IS_START(p, "; Attributes:"))
9849 {
9850 static const char *attrs[] = {
9851 "bp-based frame",
9852 "library function",
9853 "static",
9854 "noreturn",
9855 "thunk",
9856 "fpd=",
9857 };
9858
9859 // parse IDA's attribute-list comment
9860 g_ida_func_attr = 0;
9861 p = sskip(p + 13);
9862
9863 for (; *p != 0; p = sskip(p)) {
9864 for (i = 0; i < ARRAY_SIZE(attrs); i++) {
9865 if (!strncmp(p, attrs[i], strlen(attrs[i]))) {
9866 g_ida_func_attr |= 1 << i;
9867 p += strlen(attrs[i]);
9868 break;
9869 }
9870 }
9871 if (i == ARRAY_SIZE(attrs)) {
9872 anote("unparsed IDA attr: %s\n", p);
9873 break;
9874 }
9875 if (IS(attrs[i], "fpd=")) {
9876 p = next_word(words[0], sizeof(words[0]), p);
9877 // ignore for now..
9878 }
9879 }
9880 }
9881 else if (p[2] == 's' && IS_START(p, "; sctattr:"))
9882 {
9883 static const char *attrs[] = {
9884 "clear_sf",
9885 "clear_regmask",
9886 "rm_regmask",
9887 "nowarn",
9888 "argframe",
9889 "align_float",
9890 };
9891
9892 // parse manual attribute-list comment
9893 g_sct_func_attr = 0;
9894 p = sskip(p + 10);
9895
9896 for (; *p != 0; p = sskip(p)) {
9897 for (i = 0; i < ARRAY_SIZE(attrs); i++) {
9898 if (!strncmp(p, attrs[i], strlen(attrs[i]))) {
9899 g_sct_func_attr |= 1 << i;
9900 p += strlen(attrs[i]);
9901 break;
9902 }
9903 }
9904 if (*p == '=') {
9905 j = ret = 0;
9906 if (i == 0)
9907 // clear_sf=start,len (in dwords)
9908 ret = sscanf(p, "=%d,%d%n", &g_stack_clear_start,
9909 &g_stack_clear_len, &j);
9910 else if (i == 1)
9911 // clear_regmask=<mask>
9912 ret = sscanf(p, "=%x%n", &g_regmask_init, &j) + 1;
9913 else if (i == 2)
9914 // rm_regmask=<mask>
9915 ret = sscanf(p, "=%x%n", &g_regmask_rm, &j) + 1;
9916 if (ret < 2) {
9917 anote("unparsed attr value: %s\n", p);
9918 break;
9919 }
9920 p += j;
9921 }
9922 else if (i == ARRAY_SIZE(attrs)) {
9923 anote("unparsed sct attr: %s\n", p);
9924 break;
9925 }
9926 }
9927 }
9928 else if (p[2] == 'S' && IS_START(p, "; START OF FUNCTION CHUNK FOR "))
9929 {
9930 p += 30;
9931 next_word(words[0], sizeof(words[0]), p);
9932 if (words[0][0] == 0)
9933 aerr("missing name for func chunk?\n");
9934
9935 if (!scanned_ahead) {
9936 add_func_chunk(fasm, words[0], asmln);
9937 func_chunks_sorted = 0;
9938 }
9939 }
9940 else if (p[2] == 'E' && IS_START(p, "; END OF FUNCTION CHUNK"))
9941 {
9942 if (func_chunk_i >= 0) {
9943 if (func_chunk_i < func_chunk_cnt
9944 && IS(func_chunks[func_chunk_i].name, g_func))
9945 {
9946 // move on to next chunk
9947 ret = fseek(fasm, func_chunks[func_chunk_i].fptr, SEEK_SET);
9948 if (ret)
9949 aerr("seek failed for '%s' chunk #%d\n",
9950 g_func, func_chunk_i);
9951 asmln = func_chunks[func_chunk_i].asmln;
9952 func_chunk_i++;
9953 }
9954 else {
9955 if (func_chunk_ret == 0)
9956 aerr("no return from chunk?\n");
9957 fseek(fasm, func_chunk_ret, SEEK_SET);
9958 asmln = func_chunk_ret_ln;
9959 func_chunk_ret = 0;
9960 pending_endp = 1;
9961 }
9962 }
9963 }
9964 else if (p[2] == 'F' && IS_START(p, "; FUNCTION CHUNK AT ")) {
9965 func_chunks_used = 1;
9966 p += 20;
9967 if (IS_START(g_func, "sub_")) {
9968 unsigned long addr = strtoul(p, NULL, 16);
9969 unsigned long f_addr = strtoul(g_func + 4, NULL, 16);
9970 if (addr > f_addr && !scanned_ahead) {
9971 //anote("scan_ahead caused by '%s', addr %lx\n",
9972 // g_func, addr);
9973 scan_ahead_for_chunks(fasm);
9974 scanned_ahead = 1;
9975 func_chunks_sorted = 0;
9976 }
9977 }
9978 }
9979 continue;
9980 } // *p == ';'
9981
9982parse_words:
9983 for (i = wordc; i < ARRAY_SIZE(words); i++)
9984 words[i][0] = 0;
9985 for (wordc = 0; wordc < ARRAY_SIZE(words); wordc++) {
9986 p = sskip(next_word_s(words[wordc], sizeof(words[0]), p));
9987 if (*p == 0 || *p == ';') {
9988 wordc++;
9989 break;
9990 }
9991 }
9992 if (*p != 0 && *p != ';')
9993 aerr("too many words\n");
9994
9995 if (skip_code_end) {
9996 skip_code_end = 0;
9997 skip_code = 0;
9998 }
9999
10000 // allow asm patches in comments
10001 if (*p == ';') {
10002 // skip IDA's forced non-removable comment
10003 if (!IS_START(p, "; sct") && (p2 = strchr(p + 1, ';')))
10004 p = p2;
10005 }
10006 if (*p == ';' && IS_START(p, "; sct")) {
10007 if (IS_START(p, "; sctpatch:")) {
10008 p = sskip(p + 11);
10009 if (*p == 0 || *p == ';')
10010 continue;
10011 goto parse_words; // lame
10012 }
10013 else if (IS_START(p, "; sctend")) {
10014 end = 1;
10015 if (!pending_endp)
10016 break;
10017 }
10018 else if (g_skip_func)
10019 /* ignore remaining attrs */;
10020 else if (IS_START(p, "; sctproto:")) {
10021 sctproto = strdup(p + 11);
10022 }
10023 else if (IS_START(p, "; sctskip_start")) {
10024 if (in_func) {
10025 if (!skip_code) {
10026 ops[pi].op = OPP_ABORT;
10027 ops[pi].asmln = asmln;
10028 pi++;
10029 }
10030 skip_code = 1;
10031 }
10032 }
10033 else if (IS_START(p, "; sctskip_end")) {
10034 if (skip_code)
10035 skip_code_end = 1;
10036 }
10037 }
10038
10039 if (wordc == 0) {
10040 // shouldn't happen
10041 awarn("wordc == 0?\n");
10042 continue;
10043 }
10044
10045 // don't care about this:
10046 if (words[0][0] == '.'
10047 || IS(words[0], "include")
10048 || IS(words[0], "assume") || IS(words[1], "segment")
10049 || IS(words[0], "align"))
10050 {
10051 continue;
10052 }
10053
10054do_pending_endp:
10055 // do delayed endp processing to collect switch jumptables
10056 if (pending_endp) {
10057 if (in_func && !g_skip_func && !end && wordc >= 2
10058 && ((words[0][0] == 'd' && words[0][2] == 0)
10059 || (words[1][0] == 'd' && words[1][2] == 0)))
10060 {
10061 i = 1;
10062 if (words[1][0] == 'd' && words[1][2] == 0) {
10063 // label
10064 if (g_func_pd_cnt >= pd_alloc) {
10065 pd_alloc = pd_alloc * 2 + 16;
10066 g_func_pd = realloc(g_func_pd,
10067 sizeof(g_func_pd[0]) * pd_alloc);
10068 my_assert_not(g_func_pd, NULL);
10069 }
10070 pd = &g_func_pd[g_func_pd_cnt];
10071 g_func_pd_cnt++;
10072 memset(pd, 0, sizeof(*pd));
10073 strcpy(pd->label, words[0]);
10074 pd->type = OPT_CONST;
10075 pd->lmod = lmod_from_directive(words[1]);
10076 i = 2;
10077 }
10078 else {
10079 if (pd == NULL) {
10080 if (verbose)
10081 anote("skipping alignment byte?\n");
10082 continue;
10083 }
10084 lmod = lmod_from_directive(words[0]);
10085 if (lmod != pd->lmod)
10086 aerr("lmod change? %d->%d\n", pd->lmod, lmod);
10087 }
10088
10089 if (pd->count_alloc < pd->count + wordc) {
10090 pd->count_alloc = pd->count_alloc * 2 + 14 + wordc;
10091 pd->d = realloc(pd->d, sizeof(pd->d[0]) * pd->count_alloc);
10092 my_assert_not(pd->d, NULL);
10093 }
10094 for (; i < wordc; i++) {
10095 if (IS(words[i], "offset")) {
10096 pd->type = OPT_OFFSET;
10097 i++;
10098 }
10099 p = strchr(words[i], ',');
10100 if (p != NULL)
10101 *p = 0;
10102 if (pd->type == OPT_OFFSET)
10103 pd->d[pd->count].u.label = strdup(words[i]);
10104 else
10105 pd->d[pd->count].u.val = parse_number(words[i], 0);
10106 pd->d[pd->count].bt_i = -1;
10107 pd->count++;
10108 }
10109 continue;
10110 }
10111
10112 if (in_func && !g_skip_func) {
10113 if (g_header_mode)
10114 gen_hdr(g_func, pi);
10115 else
10116 gen_func(fout, g_fhdr, g_func, pi);
10117 }
10118
10119 pending_endp = 0;
10120 in_func = 0;
10121 g_ida_func_attr = 0;
10122 g_sct_func_attr = 0;
10123 g_stack_clear_start = 0;
10124 g_stack_clear_len = 0;
10125 g_regmask_init = 0;
10126 g_regmask_rm = 0;
10127 skip_warned = 0;
10128 g_skip_func = 0;
10129 g_func[0] = 0;
10130 g_seh_found = 0;
10131 func_chunks_used = 0;
10132 func_chunk_i = -1;
10133 if (pi != 0) {
10134 memset(&ops, 0, pi * sizeof(ops[0]));
10135 clear_labels(pi);
10136 pi = 0;
10137 }
10138 g_eqcnt = 0;
10139 for (i = 0; i < g_func_pd_cnt; i++) {
10140 pd = &g_func_pd[i];
10141 if (pd->type == OPT_OFFSET) {
10142 for (j = 0; j < pd->count; j++)
10143 free(pd->d[j].u.label);
10144 }
10145 free(pd->d);
10146 pd->d = NULL;
10147 }
10148 g_func_pd_cnt = 0;
10149 g_func_lmods = 0;
10150 pd = NULL;
10151
10152 if (end)
10153 break;
10154 if (wordc == 0)
10155 continue;
10156 }
10157
10158 if (IS(words[1], "proc")) {
10159 if (in_func)
10160 aerr("proc '%s' while in_func '%s'?\n",
10161 words[0], g_func);
10162 p = words[0];
10163 if (bsearch(&p, rlist, rlist_len, sizeof(rlist[0]), cmpstringp))
10164 g_skip_func = 1;
10165 strcpy(g_func, words[0]);
10166 set_label(0, words[0]);
10167 in_func = 1;
10168 continue;
10169 }
10170
10171 if (IS(words[1], "endp"))
10172 {
10173 if (!in_func)
10174 aerr("endp '%s' while not in_func?\n", words[0]);
10175 if (!IS(g_func, words[0]))
10176 aerr("endp '%s' while in_func '%s'?\n",
10177 words[0], g_func);
10178 if (skip_code)
10179 aerr("endp '%s' while skipping code\n", words[0]);
10180
10181 if ((g_ida_func_attr & IDAFA_THUNK) && pi == 1
10182 && ops[0].op == OP_JMP && ops[0].operand[0].segment)
10183 {
10184 // import jump
10185 g_skip_func = 1;
10186 }
10187
10188 if (!g_skip_func && func_chunks_used) {
10189 // start processing chunks
10190 struct chunk_item *ci, key = { g_func, 0 };
10191
10192 func_chunk_ret = ftell(fasm);
10193 func_chunk_ret_ln = asmln;
10194 if (!func_chunks_sorted) {
10195 qsort(func_chunks, func_chunk_cnt,
10196 sizeof(func_chunks[0]), cmp_chunks);
10197 func_chunks_sorted = 1;
10198 }
10199 ci = bsearch(&key, func_chunks, func_chunk_cnt,
10200 sizeof(func_chunks[0]), cmp_chunks);
10201 if (ci == NULL)
10202 aerr("'%s' needs chunks, but none found\n", g_func);
10203 func_chunk_i = ci - func_chunks;
10204 for (; func_chunk_i > 0; func_chunk_i--)
10205 if (!IS(func_chunks[func_chunk_i - 1].name, g_func))
10206 break;
10207
10208 ret = fseek(fasm, func_chunks[func_chunk_i].fptr, SEEK_SET);
10209 if (ret)
10210 aerr("seek failed for '%s' chunk #%d\n", g_func, func_chunk_i);
10211 asmln = func_chunks[func_chunk_i].asmln;
10212 func_chunk_i++;
10213 continue;
10214 }
10215 pending_endp = 1;
10216 continue;
10217 }
10218
10219 if (wordc == 2 && IS(words[1], "ends")) {
10220 if (!multi_seg) {
10221 end = 1;
10222 if (pending_endp)
10223 goto do_pending_endp;
10224 break;
10225 }
10226
10227 // scan for next text segment
10228 while (my_fgets(line, sizeof(line), fasm)) {
10229 asmln++;
10230 p = sskip(line);
10231 if (*p == 0 || *p == ';')
10232 continue;
10233
10234 if (strstr(p, "segment para public 'CODE' use32"))
10235 break;
10236 }
10237
10238 continue;
10239 }
10240
10241 p = strchr(words[0], ':');
10242 if (p != NULL) {
10243 set_label(pi, words[0]);
10244 continue;
10245 }
10246
10247 if (!in_func || g_skip_func || skip_code) {
10248 if (!skip_warned && !g_skip_func && g_labels[pi] != NULL) {
10249 if (verbose)
10250 anote("skipping from '%s'\n", g_labels[pi]);
10251 skip_warned = 1;
10252 }
10253 free(g_labels[pi]);
10254 g_labels[pi] = NULL;
10255 continue;
10256 }
10257
10258 if (wordc > 1 && IS(words[1], "="))
10259 {
10260 if (wordc != 5)
10261 aerr("unhandled equ, wc=%d\n", wordc);
10262 if (g_eqcnt >= eq_alloc) {
10263 eq_alloc *= 2;
10264 g_eqs = realloc(g_eqs, eq_alloc * sizeof(g_eqs[0]));
10265 my_assert_not(g_eqs, NULL);
10266 }
10267
10268 len = strlen(words[0]);
10269 if (len > sizeof(g_eqs[0].name) - 1)
10270 aerr("equ name too long: %d\n", len);
10271 strcpy(g_eqs[g_eqcnt].name, words[0]);
10272
10273 if (!IS(words[3], "ptr"))
10274 aerr("unhandled equ\n");
10275 if (IS(words[2], "dword"))
10276 g_eqs[g_eqcnt].lmod = OPLM_DWORD;
10277 else if (IS(words[2], "word"))
10278 g_eqs[g_eqcnt].lmod = OPLM_WORD;
10279 else if (IS(words[2], "byte"))
10280 g_eqs[g_eqcnt].lmod = OPLM_BYTE;
10281 else if (IS(words[2], "qword"))
10282 g_eqs[g_eqcnt].lmod = OPLM_QWORD;
10283 else
10284 aerr("bad lmod: '%s'\n", words[2]);
10285
10286 g_eqs[g_eqcnt].offset = parse_number(words[4], 0);
10287 g_eqcnt++;
10288 continue;
10289 }
10290
10291 if (pi >= ARRAY_SIZE(ops))
10292 aerr("too many ops\n");
10293
10294 parse_op(&ops[pi], words, wordc);
10295
10296 ops[pi].datap = sctproto;
10297 sctproto = NULL;
10298 pi++;
10299 }
10300
10301 if (g_header_mode)
10302 output_hdr(fout);
10303
10304 fclose(fout);
10305 fclose(fasm);
10306 fclose(g_fhdr);
10307
10308 return 0;
10309}
10310
10311// vim:ts=2:shiftwidth=2:expandtab
Unnamed repository; edit this file 'description' to name the repository.