X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=tools%2Ftranslate.c;h=1fd5e4715b5e276f467070356d13c74bcd181901;hb=5c024ef78bfc9bac8b67870c169ab8732d8f7d89;hp=9643c63f2c6d34445290fedbf6047167b68e4625;hpb=ddaf8bd7aa44ee867e43988ba0bb7658dae6e6eb;p=ia32rtools.git diff --git a/tools/translate.c b/tools/translate.c index 9643c63..1fd5e47 100644 --- a/tools/translate.c +++ b/tools/translate.c @@ -32,15 +32,17 @@ enum op_flags { OPF_RMD = (1 << 0), /* removed or optimized out */ OPF_DATA = (1 << 1), /* data processing - writes to dst opr */ OPF_FLAGS = (1 << 2), /* sets flags */ - OPF_JMP = (1 << 3), /* branches, ret and call */ - OPF_CC = (1 << 4), /* uses flags */ - OPF_TAIL = (1 << 5), /* ret or tail call */ - OPF_RSAVE = (1 << 6), /* push/pop is local reg save/load */ - OPF_REP = (1 << 7), /* prefixed by rep */ - OPF_REPZ = (1 << 8), /* rep is repe/repz */ - OPF_REPNZ = (1 << 9), /* rep is repne/repnz */ - OPF_FARG = (1 << 10), /* push collected as func arg (no reuse) */ - OPF_EBP_S = (1 << 11), /* ebp used as scratch here, not BP */ + OPF_JMP = (1 << 3), /* branch, call */ + OPF_CJMP = (1 << 4), /* cond. branch (cc or jecxz) */ + OPF_CC = (1 << 5), /* uses flags */ + OPF_TAIL = (1 << 6), /* ret or tail call */ + OPF_RSAVE = (1 << 7), /* push/pop is local reg save/load */ + OPF_REP = (1 << 8), /* prefixed by rep */ + OPF_REPZ = (1 << 9), /* rep is repe/repz */ + OPF_REPNZ = (1 << 10), /* rep is repne/repnz */ + OPF_FARG = (1 << 11), /* push collected as func arg (no reuse) */ + OPF_EBP_S = (1 << 12), /* ebp used as scratch here, not BP */ + OPF_DF = (1 << 13), /* DF flag set */ }; enum op_op { @@ -48,6 +50,7 @@ enum op_op { OP_NOP, OP_PUSH, OP_POP, + OP_LEAVE, OP_MOV, OP_LEA, OP_MOVZX, @@ -57,6 +60,9 @@ enum op_op { OP_STOS, OP_MOVS, OP_CMPS, + OP_SCAS, + OP_STD, + OP_CLD, OP_RET, OP_ADD, OP_SUB, @@ -81,6 +87,7 @@ enum op_op { OP_CMP, OP_CALL, OP_JMP, + OP_JECXZ, OP_JO, OP_JNO, OP_JC, @@ -150,6 +157,7 @@ struct parsed_op { // datap: // OP_CALL - parser proto hint (str), ptr to struct parsed_proto // (OPF_CC) - point to one of (OPF_FLAGS) that affects cc op +// OP_POP - point to OP_PUSH in push/pop pair struct parsed_equ { char name[64]; @@ -398,7 +406,7 @@ static const char *parse_stack_el(const char *name, char *extra_reg) // must be a number after esp+, already converted to 0x.. s = name + 4; if (!('0' <= *s && *s <= '9')) { - aerr("%s nan?\n", __func__); + aerr("%s nan?\n", __func__); return NULL; } if (s[0] == '0' && s[1] == 'x') @@ -701,6 +709,8 @@ static const struct { { "repnz", OPF_REP|OPF_REPNZ }, }; +#define OPF_CJMP_CC (OPF_JMP|OPF_CJMP|OPF_CC) + static const struct { const char *name; enum op_op op; @@ -711,6 +721,7 @@ static const struct { { "nop", OP_NOP, 0, 0, 0 }, { "push", OP_PUSH, 1, 1, 0 }, { "pop", OP_POP, 1, 1, OPF_DATA }, + { "leave",OP_LEAVE, 0, 0, OPF_DATA }, { "mov" , OP_MOV, 2, 2, OPF_DATA }, { "lea", OP_LEA, 2, 2, OPF_DATA }, { "movzx",OP_MOVZX, 2, 2, OPF_DATA }, @@ -726,6 +737,11 @@ static const struct { { "cmpsb",OP_CMPS, 0, 0, OPF_DATA|OPF_FLAGS }, { "cmpsw",OP_CMPS, 0, 0, OPF_DATA|OPF_FLAGS }, { "cmpsd",OP_CMPS, 0, 0, OPF_DATA|OPF_FLAGS }, + { "scasb",OP_SCAS, 0, 0, OPF_DATA|OPF_FLAGS }, + { "scasw",OP_SCAS, 0, 0, OPF_DATA|OPF_FLAGS }, + { "scasd",OP_SCAS, 0, 0, OPF_DATA|OPF_FLAGS }, + { "std", OP_STD, 0, 0, OPF_DATA }, // special flag + { "cld", OP_CLD, 0, 0, OPF_DATA }, { "add", OP_ADD, 2, 2, OPF_DATA|OPF_FLAGS }, { "sub", OP_SUB, 2, 2, OPF_DATA|OPF_FLAGS }, { "and", OP_AND, 2, 2, OPF_DATA|OPF_FLAGS }, @@ -751,35 +767,36 @@ static const struct { { "retn", OP_RET, 0, 1, OPF_TAIL }, { "call", OP_CALL, 1, 1, OPF_JMP|OPF_DATA|OPF_FLAGS }, { "jmp", OP_JMP, 1, 1, OPF_JMP }, - { "jo", OP_JO, 1, 1, OPF_JMP|OPF_CC }, // 70 OF=1 - { "jno", OP_JNO, 1, 1, OPF_JMP|OPF_CC }, // 71 OF=0 - { "jc", OP_JC, 1, 1, OPF_JMP|OPF_CC }, // 72 CF=1 - { "jb", OP_JC, 1, 1, OPF_JMP|OPF_CC }, // 72 - { "jnc", OP_JNC, 1, 1, OPF_JMP|OPF_CC }, // 73 CF=0 - { "jnb", OP_JNC, 1, 1, OPF_JMP|OPF_CC }, // 73 - { "jae", OP_JNC, 1, 1, OPF_JMP|OPF_CC }, // 73 - { "jz", OP_JZ, 1, 1, OPF_JMP|OPF_CC }, // 74 ZF=1 - { "je", OP_JZ, 1, 1, OPF_JMP|OPF_CC }, // 74 - { "jnz", OP_JNZ, 1, 1, OPF_JMP|OPF_CC }, // 75 ZF=0 - { "jne", OP_JNZ, 1, 1, OPF_JMP|OPF_CC }, // 75 - { "jbe", OP_JBE, 1, 1, OPF_JMP|OPF_CC }, // 76 CF=1 || ZF=1 - { "jna", OP_JBE, 1, 1, OPF_JMP|OPF_CC }, // 76 - { "ja", OP_JA, 1, 1, OPF_JMP|OPF_CC }, // 77 CF=0 && ZF=0 - { "jnbe", OP_JA, 1, 1, OPF_JMP|OPF_CC }, // 77 - { "js", OP_JS, 1, 1, OPF_JMP|OPF_CC }, // 78 SF=1 - { "jns", OP_JNS, 1, 1, OPF_JMP|OPF_CC }, // 79 SF=0 - { "jp", OP_JP, 1, 1, OPF_JMP|OPF_CC }, // 7a PF=1 - { "jpe", OP_JP, 1, 1, OPF_JMP|OPF_CC }, // 7a - { "jnp", OP_JNP, 1, 1, OPF_JMP|OPF_CC }, // 7b PF=0 - { "jpo", OP_JNP, 1, 1, OPF_JMP|OPF_CC }, // 7b - { "jl", OP_JL, 1, 1, OPF_JMP|OPF_CC }, // 7c SF!=OF - { "jnge", OP_JL, 1, 1, OPF_JMP|OPF_CC }, // 7c - { "jge", OP_JGE, 1, 1, OPF_JMP|OPF_CC }, // 7d SF=OF - { "jnl", OP_JGE, 1, 1, OPF_JMP|OPF_CC }, // 7d - { "jle", OP_JLE, 1, 1, OPF_JMP|OPF_CC }, // 7e ZF=1 || SF!=OF - { "jng", OP_JLE, 1, 1, OPF_JMP|OPF_CC }, // 7e - { "jg", OP_JG, 1, 1, OPF_JMP|OPF_CC }, // 7f ZF=0 && SF=OF - { "jnle", OP_JG, 1, 1, OPF_JMP|OPF_CC }, // 7f + { "jecxz",OP_JECXZ, 1, 1, OPF_JMP|OPF_CJMP }, + { "jo", OP_JO, 1, 1, OPF_CJMP_CC }, // 70 OF=1 + { "jno", OP_JNO, 1, 1, OPF_CJMP_CC }, // 71 OF=0 + { "jc", OP_JC, 1, 1, OPF_CJMP_CC }, // 72 CF=1 + { "jb", OP_JC, 1, 1, OPF_CJMP_CC }, // 72 + { "jnc", OP_JNC, 1, 1, OPF_CJMP_CC }, // 73 CF=0 + { "jnb", OP_JNC, 1, 1, OPF_CJMP_CC }, // 73 + { "jae", OP_JNC, 1, 1, OPF_CJMP_CC }, // 73 + { "jz", OP_JZ, 1, 1, OPF_CJMP_CC }, // 74 ZF=1 + { "je", OP_JZ, 1, 1, OPF_CJMP_CC }, // 74 + { "jnz", OP_JNZ, 1, 1, OPF_CJMP_CC }, // 75 ZF=0 + { "jne", OP_JNZ, 1, 1, OPF_CJMP_CC }, // 75 + { "jbe", OP_JBE, 1, 1, OPF_CJMP_CC }, // 76 CF=1 || ZF=1 + { "jna", OP_JBE, 1, 1, OPF_CJMP_CC }, // 76 + { "ja", OP_JA, 1, 1, OPF_CJMP_CC }, // 77 CF=0 && ZF=0 + { "jnbe", OP_JA, 1, 1, OPF_CJMP_CC }, // 77 + { "js", OP_JS, 1, 1, OPF_CJMP_CC }, // 78 SF=1 + { "jns", OP_JNS, 1, 1, OPF_CJMP_CC }, // 79 SF=0 + { "jp", OP_JP, 1, 1, OPF_CJMP_CC }, // 7a PF=1 + { "jpe", OP_JP, 1, 1, OPF_CJMP_CC }, // 7a + { "jnp", OP_JNP, 1, 1, OPF_CJMP_CC }, // 7b PF=0 + { "jpo", OP_JNP, 1, 1, OPF_CJMP_CC }, // 7b + { "jl", OP_JL, 1, 1, OPF_CJMP_CC }, // 7c SF!=OF + { "jnge", OP_JL, 1, 1, OPF_CJMP_CC }, // 7c + { "jge", OP_JGE, 1, 1, OPF_CJMP_CC }, // 7d SF=OF + { "jnl", OP_JGE, 1, 1, OPF_CJMP_CC }, // 7d + { "jle", OP_JLE, 1, 1, OPF_CJMP_CC }, // 7e ZF=1 || SF!=OF + { "jng", OP_JLE, 1, 1, OPF_CJMP_CC }, // 7e + { "jg", OP_JG, 1, 1, OPF_CJMP_CC }, // 7f ZF=0 && SF=OF + { "jnle", OP_JG, 1, 1, OPF_CJMP_CC }, // 7f { "seto", OP_JO, 1, 1, OPF_DATA|OPF_CC }, { "setno", OP_JNO, 1, 1, OPF_DATA|OPF_CC }, { "setc", OP_JC, 1, 1, OPF_DATA|OPF_CC }, @@ -883,13 +900,14 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc) break; case OP_STOS: + case OP_SCAS: if (op->operand_cnt != 0) break; - if (IS(words[op_w], "stosb")) + if (words[op_w][4] == 'b') lmod = OPLM_BYTE; - else if (IS(words[op_w], "stosw")) + else if (words[op_w][4] == 'w') lmod = OPLM_WORD; - else if (IS(words[op_w], "stosd")) + else if (words[op_w][4] == 'd') lmod = OPLM_DWORD; op->operand_cnt = 3; setup_reg_opr(&op->operand[0], xDI, lmod, &op->regmask_src); @@ -915,6 +933,14 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc) op->regmask_dst = op->regmask_src; break; + case OP_JECXZ: + op->operand_cnt = 1; + op->regmask_src = 1 << xCX; + op->operand[0].type = OPT_REG; + op->operand[0].reg = xCX; + op->operand[0].lmod = OPLM_DWORD; + break; + case OP_IMUL: if (op->operand_cnt != 1) break; @@ -1450,9 +1476,13 @@ static void stack_frame_access(struct parsed_op *po, static void check_func_pp(struct parsed_op *po, const struct parsed_proto *pp, const char *pfx) { + char buf[256]; + if (pp->argc_reg != 0) { - if (!g_allow_regfunc) - ferr(po, "%s: reg arg in arg-call unhandled yet\n", pfx); + if (/*!g_allow_regfunc &&*/ !pp->is_fastcall) { + pp_print(buf, sizeof(buf), pp); + ferr(po, "%s: unexpected reg arg in icall: %s\n", pfx, buf); + } if (pp->argc_stack > 0 && pp->argc_reg != 2) ferr(po, "%s: %d reg arg(s) with %d stack arg(s)\n", pfx, pp->argc_reg, pp->argc_stack); @@ -1472,7 +1502,7 @@ static void check_label_read_ref(struct parsed_op *po, const char *name) } static char *out_src_opr(char *buf, size_t buf_size, - struct parsed_op *po, struct parsed_opr *popr, const char *cast, + struct parsed_op *po, struct parsed_opr *popr, const char *cast, int is_lea) { char tmp1[256], tmp2[256]; @@ -1909,7 +1939,7 @@ static void op_set_clear_flag(struct parsed_op *po, // last op in stream - unconditional branch or ret #define LAST_OP(_i) ((ops[_i].flags & OPF_TAIL) \ - || (ops[_i].flags & (OPF_JMP|OPF_CC)) == OPF_JMP) + || (ops[_i].flags & (OPF_JMP|OPF_CJMP)) == OPF_JMP) static int scan_for_pop(int i, int opcnt, const char *reg, int magic, int depth, int *maxdepth, int do_flags) @@ -1956,7 +1986,7 @@ static int scan_for_pop(int i, int opcnt, const char *reg, return -1; } - if (po->flags & OPF_CC) { + if (po->flags & OPF_CJMP) { ret |= scan_for_pop(po->bt_i, opcnt, reg, magic, depth, maxdepth, do_flags); if (ret < 0) @@ -2032,6 +2062,52 @@ static int scan_for_pop_ret(int i, int opcnt, const char *reg, return found ? 0 : -1; } +static void scan_propagate_df(int i, int opcnt) +{ + struct parsed_op *po = &ops[i]; + int j; + + for (; i < opcnt; i++) { + po = &ops[i]; + if (po->flags & OPF_DF) + return; // already resolved + po->flags |= OPF_DF; + + if (po->op == OP_CALL) + ferr(po, "call with DF set?\n"); + + if (po->flags & OPF_JMP) { + if (po->btj != NULL) { + // jumptable + for (j = 0; j < po->btj->count; j++) + scan_propagate_df(po->btj->d[j].bt_i, opcnt); + return; + } + + if (po->bt_i < 0) { + ferr(po, "dead branch\n"); + return; + } + + if (po->flags & OPF_CJMP) + scan_propagate_df(po->bt_i, opcnt); + else + i = po->bt_i - 1; + continue; + } + + if (po->flags & OPF_TAIL) + break; + + if (po->op == OP_CLD) { + po->flags |= OPF_RMD; + return; + } + } + + ferr(po, "missing DF clear?\n"); +} + // is operand 'opr' modified by parsed_op 'po'? static int is_opr_modified(const struct parsed_opr *opr, const struct parsed_op *po) @@ -2162,7 +2238,7 @@ static int scan_for_flag_set(int i, int magic, int *branched, return 0; } - if ((ops[i].flags & OPF_JMP) && !(ops[i].flags & OPF_CC)) + if ((ops[i].flags & (OPF_JMP|OPF_CJMP)) == OPF_JMP) return -1; } @@ -2244,12 +2320,12 @@ static int scan_for_esp_adjust(int i, int opcnt, int *adj) return i; } else if (po->op == OP_PUSH) { - if (first_pop == -1) - first_pop = -2; // none + //if (first_pop == -1) + // first_pop = -2; // none *adj -= lmod_bytes(po, po->operand[0].lmod); } - else if (po->op == OP_POP) { - if (first_pop == -1) + else if (po->op == OP_POP && !(po->flags & OPF_RMD)) { + if (first_pop == -1 && *adj >= 0) first_pop = i; *adj += lmod_bytes(po, po->operand[0].lmod); } @@ -2264,7 +2340,7 @@ static int scan_for_esp_adjust(int i, int opcnt, int *adj) } } - if (*adj == 4 && first_pop >= 0 && ops[first_pop].op == OP_POP + if (*adj <= 8 && first_pop >= 0 && ops[first_pop].op == OP_POP && ops[first_pop].operand[0].type == OPT_REG && ops[first_pop].operand[0].reg == xCX) { @@ -2299,7 +2375,7 @@ static void scan_fwd_set_flags(int i, int opcnt, int magic, int flags) } scan_fwd_set_flags(po->bt_i, opcnt, magic, flags); - if (!(po->flags & OPF_CC)) + if (!(po->flags & OPF_CJMP)) return; } if (po->flags & OPF_TAIL) @@ -2505,7 +2581,7 @@ static int collect_call_args_r(struct parsed_op *po, int i, if (lr->next != NULL) need_op_saving = 1; for (; lr->next; lr = lr->next) { - if ((ops[lr->i].flags & (OPF_JMP|OPF_CC)) != OPF_JMP) + if ((ops[lr->i].flags & (OPF_JMP|OPF_CJMP)) != OPF_JMP) may_reuse = 1; ret = collect_call_args_r(po, lr->i, pp, regmask, save_arg_vars, arg, magic, need_op_saving, may_reuse); @@ -2513,7 +2589,7 @@ static int collect_call_args_r(struct parsed_op *po, int i, return ret; } - if ((ops[lr->i].flags & (OPF_JMP|OPF_CC)) != OPF_JMP) + if ((ops[lr->i].flags & (OPF_JMP|OPF_CJMP)) != OPF_JMP) may_reuse = 1; if (j > 0 && LAST_OP(j - 1)) { // follow last branch in reverse @@ -2554,7 +2630,7 @@ static int collect_call_args_r(struct parsed_op *po, int i, ferr(po, "arg collect %d/%d hit pop\n", arg, pp->argc); } - else if ((ops[j].flags & (OPF_JMP|OPF_CC)) == (OPF_JMP|OPF_CC)) + else if (ops[j].flags & OPF_CJMP) { if (pp->is_unresolved) break; @@ -2689,6 +2765,17 @@ static void output_std_flags(FILE *fout, struct parsed_op *po, } } +static void output_pp_attrs(FILE *fout, const struct parsed_proto *pp, + int is_noreturn) +{ + if (pp->is_fastcall) + fprintf(fout, "__fastcall "); + else if (pp->is_stdcall && pp->argc_reg == 0) + fprintf(fout, "__stdcall "); + if (pp->is_noreturn || is_noreturn) + fprintf(fout, "noreturn "); +} + static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) { struct parsed_op *po, *delayed_flag_op = NULL, *tmp_op; @@ -2727,10 +2814,7 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) ferr(ops, "proto_parse failed for '%s'\n", funcn); fprintf(fout, "%s ", g_func_pp->ret_type.name); - if (g_func_pp->is_stdcall && g_func_pp->argc_reg == 0) - fprintf(fout, "__stdcall "); - if (g_ida_func_attr & IDAFA_NORETURN) - fprintf(fout, "noreturn "); + output_pp_attrs(fout, g_func_pp, g_ida_func_attr & IDAFA_NORETURN); fprintf(fout, "%s(", funcn); for (i = 0; i < g_func_pp->argc; i++) { @@ -2740,8 +2824,7 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) // func pointer.. pp = g_func_pp->arg[i].fptr; fprintf(fout, "%s (", pp->ret_type.name); - if (pp->is_stdcall && pp->argc_reg == 0) - fprintf(fout, "__stdcall "); + output_pp_attrs(fout, pp, 0); fprintf(fout, "*a%d)(", i + 1); for (j = 0; j < pp->argc; j++) { if (j > 0) @@ -2821,8 +2904,11 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) if (i == opcnt && (ops[i - 1].flags & OPF_JMP) && found) break; - if (ops[i - 1].op == OP_POP && IS(opr_name(&ops[i - 1], 0), "ebp")) + if ((ops[i - 1].op == OP_POP && IS(opr_name(&ops[i - 1], 0), "ebp")) + || ops[i - 1].op == OP_LEAVE) + { ops[i - 1].flags |= OPF_RMD; + } else if (!(g_ida_func_attr & IDAFA_NORETURN)) ferr(&ops[i - 1], "'pop ebp' expected\n"); @@ -2833,8 +2919,11 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) { ops[i - 2].flags |= OPF_RMD; } - else if (!(g_ida_func_attr & IDAFA_NORETURN)) + else if (ops[i - 1].op != OP_LEAVE + && !(g_ida_func_attr & IDAFA_NORETURN)) + { ferr(&ops[i - 2], "esp restore expected\n"); + } if (ecx_push && ops[i - 3].op == OP_POP && IS(opr_name(&ops[i - 3], 0), "ecx")) @@ -3062,9 +3151,20 @@ tailcall: tmpname, pp->argc_stack * 4, j); ops[ret].flags |= OPF_RMD; - // a bit of a hack, but deals with use of - // single adj for multiple calls - ops[ret].operand[1].val -= j; + if (ops[ret].op == OP_POP && j > 4) { + // deal with multi-pop stack adjust + j = pp->argc_stack; + while (ops[ret].op == OP_POP && j > 0 && ret < opcnt) { + ops[ret].flags |= OPF_RMD; + j--; + ret++; + } + } + else { + // a bit of a hack, but deals with use of + // single adj for multiple calls + ops[ret].operand[1].val -= j; + } } else if (pp->is_vararg) ferr(po, "missing esp_adjust for vararg func '%s'\n", @@ -3083,6 +3183,7 @@ tailcall: // pass4: // - find POPs for PUSHes, rm both + // - scan for STD/CLD, propagate DF // - scan for all used registers // - find flag set ops for their users // - do unreselved calls @@ -3101,40 +3202,63 @@ tailcall: regmask_save |= 1 << reg; } - if (po->op == OP_PUSH - && po->argnum == 0 && !(po->flags & OPF_RSAVE) - && po->operand[0].type == OPT_REG) + if (po->op == OP_PUSH && po->argnum == 0 + && !(po->flags & OPF_RSAVE)) { - reg = po->operand[0].reg; - if (reg < 0) - ferr(po, "reg not set for push?\n"); - - depth = 0; - ret = scan_for_pop(i + 1, opcnt, - po->operand[0].name, i + opcnt * 3, 0, &depth, 0); - if (ret == 1) { - if (depth > 1) - ferr(po, "too much depth: %d\n", depth); - - po->flags |= OPF_RMD; - scan_for_pop(i + 1, opcnt, po->operand[0].name, - i + opcnt * 4, 0, &depth, 1); - continue; + if (po->operand[0].type == OPT_REG) + { + reg = po->operand[0].reg; + if (reg < 0) + ferr(po, "reg not set for push?\n"); + + depth = 0; + ret = scan_for_pop(i + 1, opcnt, + po->operand[0].name, i + opcnt * 3, 0, &depth, 0); + if (ret == 1) { + if (depth > 1) + ferr(po, "too much depth: %d\n", depth); + + po->flags |= OPF_RMD; + scan_for_pop(i + 1, opcnt, po->operand[0].name, + i + opcnt * 4, 0, &depth, 1); + continue; + } + ret = scan_for_pop_ret(i + 1, opcnt, po->operand[0].name, 0); + if (ret == 0) { + arg = OPF_RMD; + if (regmask & (1 << reg)) { + if (regmask_save & (1 << reg)) + ferr(po, "%s already saved?\n", po->operand[0].name); + arg = OPF_RSAVE; + } + po->flags |= arg; + scan_for_pop_ret(i + 1, opcnt, po->operand[0].name, arg); + continue; + } } - ret = scan_for_pop_ret(i + 1, opcnt, po->operand[0].name, 0); - if (ret == 0) { - arg = OPF_RMD; - if (regmask & (1 << reg)) { - if (regmask_save & (1 << reg)) - ferr(po, "%s already saved?\n", po->operand[0].name); - arg = OPF_RSAVE; + else if (po->operand[0].type == OPT_CONST) { + for (j = i + 1; j < opcnt; j++) { + if ((ops[j].flags & (OPF_JMP|OPF_TAIL|OPF_RSAVE)) + || ops[j].op == OP_PUSH || g_labels[i][0] != 0) + { + break; + } + + if (!(ops[j].flags & OPF_RMD) && ops[j].op == OP_POP) + { + po->flags |= OPF_RMD; + ops[j].datap = po; + break; + } } - po->flags |= arg; - scan_for_pop_ret(i + 1, opcnt, po->operand[0].name, arg); - continue; } } + if (po->op == OP_STD) { + po->flags |= OPF_DF | OPF_RMD; + scan_propagate_df(i + 1, opcnt); + } + regmask_now = po->regmask_src | po->regmask_dst; if (regmask_now & (1 << xBP)) { if (g_bp_frame && !(po->flags & OPF_EBP_S)) { @@ -3167,11 +3291,15 @@ tailcall: // to get nicer code, we try to delay test and cmp; // if we can't because of operand modification, or if we - // have math op, or branch, make it calculate flags explicitly - if (tmp_op->op == OP_TEST || tmp_op->op == OP_CMP) { + // have arith op, or branch, make it calculate flags explicitly + if (tmp_op->op == OP_TEST || tmp_op->op == OP_CMP) + { if (branched || scan_for_mod(tmp_op, setters[j] + 1, i, 0) >= 0) pfomask = 1 << pfo; } + else if (tmp_op->op == OP_SCAS) { + pfomask = 1 << pfo; + } else if (tmp_op->op == OP_CMPS) { pfomask = 1 << PFO_Z; } @@ -3194,6 +3322,9 @@ tailcall: if (po->op == OP_ADC || po->op == OP_SBB) cmp_result_vars |= 1 << PFO_C; } + else if (po->op == OP_CMPS || po->op == OP_SCAS) { + cmp_result_vars |= 1 << PFO_Z; + } else if (po->op == OP_MUL || (po->op == OP_IMUL && po->operand_cnt == 1)) { @@ -3206,16 +3337,30 @@ tailcall: if (pp->is_unresolved) { int regmask_stack = 0; - collect_call_args(po, i, pp, ®mask_stack, &save_arg_vars, + collect_call_args(po, i, pp, ®mask, &save_arg_vars, i + opcnt * 2); + // this is pretty rough guess: + // see ecx and edx were pushed (and not their saved versions) + for (arg = 0; arg < pp->argc; arg++) { + if (pp->arg[arg].reg != NULL) + continue; + + tmp_op = pp->arg[arg].datap; + if (tmp_op == NULL) + ferr(po, "parsed_op missing for arg%d\n", arg); + if (tmp_op->argnum == 0 && tmp_op->operand[0].type == OPT_REG) + regmask_stack |= 1 << tmp_op->operand[0].reg; + } + if (!((regmask_stack & (1 << xCX)) && (regmask_stack & (1 << xDX)))) { if (pp->argc_stack != 0 - || ((regmask | regmask_arg) & (1 << xCX))) + || ((regmask | regmask_arg) & ((1 << xCX)|(1 << xDX)))) { pp_insert_reg_arg(pp, "ecx"); + pp->is_fastcall = 1; regmask_init |= 1 << xCX; regmask |= 1 << xCX; } @@ -3227,7 +3372,10 @@ tailcall: regmask |= 1 << xDX; } } - regmask |= regmask_stack; + + // note: __cdecl doesn't fall into is_unresolved category + if (pp->argc_stack > 0) + pp->is_stdcall = 1; } for (arg = 0; arg < pp->argc; arg++) { @@ -3265,8 +3413,7 @@ tailcall: snprintf(pp->name, sizeof(pp->name), "icall%d", i); fprintf(fout, " %s (", pp->ret_type.name); - if (pp->is_stdcall && pp->argc_reg == 0) - fprintf(fout, "__stdcall "); + output_pp_attrs(fout, pp, 0); fprintf(fout, "*%s)(", pp->name); for (j = 0; j < pp->argc; j++) { if (j > 0) @@ -3540,53 +3687,54 @@ tailcall: break; case OP_STOS: - // assumes DF=0 assert_operand_cnt(3); if (po->flags & OPF_REP) { - fprintf(fout, " for (; ecx != 0; ecx--, edi += %d)\n", + fprintf(fout, " for (; ecx != 0; ecx--, edi %c= %d)\n", + (po->flags & OPF_DF) ? '-' : '+', lmod_bytes(po, po->operand[0].lmod)); fprintf(fout, " %sedi = eax;", lmod_cast_u_ptr(po, po->operand[0].lmod)); strcpy(g_comment, "rep stos"); } else { - fprintf(fout, " %sedi = eax; edi += %d;", + fprintf(fout, " %sedi = eax; edi %c= %d;", lmod_cast_u_ptr(po, po->operand[0].lmod), + (po->flags & OPF_DF) ? '-' : '+', lmod_bytes(po, po->operand[0].lmod)); strcpy(g_comment, "stos"); } break; case OP_MOVS: - // assumes DF=0 assert_operand_cnt(3); j = lmod_bytes(po, po->operand[0].lmod); strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod)); + l = (po->flags & OPF_DF) ? '-' : '+'; if (po->flags & OPF_REP) { fprintf(fout, - " for (; ecx != 0; ecx--, edi += %d, esi += %d)\n", - j, j); + " for (; ecx != 0; ecx--, edi %c= %d, esi %c= %d)\n", + l, j, l, j); fprintf(fout, " %sedi = %sesi;", buf1, buf1); strcpy(g_comment, "rep movs"); } else { - fprintf(fout, " %sedi = %sesi; edi += %d; esi += %d;", - buf1, buf1, j, j); + fprintf(fout, " %sedi = %sesi; edi %c= %d; esi %c= %d;", + buf1, buf1, l, j, l, j); strcpy(g_comment, "movs"); } break; case OP_CMPS: - // assumes DF=0 // repe ~ repeat while ZF=1 assert_operand_cnt(3); j = lmod_bytes(po, po->operand[0].lmod); strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod)); + l = (po->flags & OPF_DF) ? '-' : '+'; if (po->flags & OPF_REP) { fprintf(fout, - " for (; ecx != 0; ecx--, edi += %d, esi += %d)\n", - j, j); + " for (; ecx != 0; ecx--, edi %c= %d, esi %c= %d)\n", + l, j, l, j); fprintf(fout, " if ((cond_z = (%sedi == %sesi)) %s 0)\n", buf1, buf1, (po->flags & OPF_REPZ) ? "==" : "!="); @@ -3597,8 +3745,8 @@ tailcall: } else { fprintf(fout, - " cond_z = (%sedi = %sesi); edi += %d; esi += %d;", - buf1, buf1, j, j); + " cond_z = (%sedi = %sesi); edi %c= %d; esi %c= %d;", + buf1, buf1, l, j, l, j); strcpy(g_comment, "cmps"); } pfomask &= ~(1 << PFO_Z); @@ -3606,6 +3754,36 @@ tailcall: delayed_flag_op = NULL; break; + case OP_SCAS: + // only does ZF (for now) + // repe ~ repeat while ZF=1 + assert_operand_cnt(3); + j = lmod_bytes(po, po->operand[0].lmod); + l = (po->flags & OPF_DF) ? '-' : '+'; + if (po->flags & OPF_REP) { + fprintf(fout, + " for (; ecx != 0; ecx--, edi %c= %d)\n", l, j); + fprintf(fout, + " if ((cond_z = (%seax == %sedi)) %s 0)\n", + lmod_cast_u(po, po->operand[0].lmod), + lmod_cast_u_ptr(po, po->operand[0].lmod), + (po->flags & OPF_REPZ) ? "==" : "!="); + fprintf(fout, + " break;"); + snprintf(g_comment, sizeof(g_comment), "rep%s scas", + (po->flags & OPF_REPZ) ? "e" : "ne"); + } + else { + fprintf(fout, " cond_z = (%seax = %sedi); edi %c= %d;", + lmod_cast_u(po, po->operand[0].lmod), + lmod_cast_u_ptr(po, po->operand[0].lmod), l, j); + strcpy(g_comment, "scas"); + } + pfomask &= ~(1 << PFO_Z); + last_arith_dst = NULL; + delayed_flag_op = NULL; + break; + // arithmetic w/flags case OP_ADD: case OP_SUB: @@ -3819,6 +3997,12 @@ tailcall: // else SETcc - should already be handled break; + case OP_JECXZ: + fprintf(fout, " if (ecx == 0)\n"); + fprintf(fout, " goto %s;", po->operand[0].name); + strcat(g_comment, "jecxz"); + break; + case OP_JMP: assert_operand_cnt(1); last_arith_dst = NULL; @@ -3985,6 +4169,16 @@ tailcall: fprintf(fout, " %s = s_%s;", buf1, buf1); break; } + else if (po->datap != NULL) { + // push/pop pair + tmp_op = po->datap; + out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]); + fprintf(fout, " %s = %s;", buf1, + out_src_opr(buf2, sizeof(buf2), + tmp_op, &tmp_op->operand[0], + po->operand[0].is_ptr ? "(void *)" : "", 0)); + break; + } ferr(po, "stray pop encountered\n"); break; @@ -4011,10 +4205,16 @@ tailcall: fprintf(fout, "\n"); // some sanity checking - if ((po->flags & OPF_REP) && po->op != OP_STOS - && po->op != OP_MOVS && po->op != OP_CMPS) - ferr(po, "unexpected rep\n"); - if ((po->flags & (OPF_REPZ|OPF_REPNZ)) && po->op != OP_CMPS) + if (po->flags & OPF_REP) { + if (po->op != OP_STOS && po->op != OP_MOVS + && po->op != OP_CMPS && po->op != OP_SCAS) + ferr(po, "unexpected rep\n"); + if (!(po->flags & (OPF_REPZ|OPF_REPNZ)) + && (po->op == OP_CMPS || po->op == OP_SCAS)) + ferr(po, "cmps/scas with plain rep\n"); + } + if ((po->flags & (OPF_REPZ|OPF_REPNZ)) + && po->op != OP_CMPS && po->op != OP_SCAS) ferr(po, "unexpected repz/repnz\n"); if (pfomask != 0)