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