+ ops[j].flags |= OPF_RMD | OPF_DONE;
+ ops[j].operand[1].val = 0; // hack for stack arg scanner
+ found = 1;
+ }
+
+ i++;
+ } while (i < opcnt);
+ }
+}
+
+static const struct parsed_proto *resolve_icall(int i, int opcnt,
+ int *multi_src)
+{
+ const struct parsed_proto *pp = NULL;
+ int search_advice = 0;
+
+ *multi_src = 0;
+
+ switch (ops[i].operand[0].type) {
+ case OPT_REGMEM:
+ case OPT_LABEL:
+ case OPT_OFFSET:
+ pp = try_recover_pp(&ops[i], &ops[i].operand[0], &search_advice);
+ if (!search_advice)
+ break;
+ // fallthrough
+ default:
+ scan_for_call_type(i, &ops[i].operand[0], i + opcnt * 9, &pp,
+ multi_src);
+ break;
+ }
+
+ return pp;
+}
+
+// find an instruction that changed opr before i op
+// *op_i must be set to -1 by caller
+// *entry is set to 1 if one source is determined to be the caller
+// returns 1 if found, *op_i is then set to origin
+static int resolve_origin(int i, const struct parsed_opr *opr,
+ int magic, int *op_i, int *is_caller)
+{
+ struct label_ref *lr;
+ int ret = 0;
+
+ if (ops[i].cc_scratch == magic)
+ return 0;
+ ops[i].cc_scratch = magic;
+
+ while (1) {
+ if (g_labels[i] != NULL) {
+ lr = &g_label_refs[i];
+ for (; lr != NULL; lr = lr->next) {
+ check_i(&ops[i], lr->i);
+ ret |= resolve_origin(lr->i, opr, magic, op_i, is_caller);
+ }
+ if (i > 0 && LAST_OP(i - 1))
+ return ret;
+ }
+
+ i--;
+ if (i < 0) {
+ if (is_caller != NULL)
+ *is_caller = 1;
+ return -1;
+ }
+
+ if (ops[i].cc_scratch == magic)
+ return 0;
+ ops[i].cc_scratch = magic;
+
+ if (!(ops[i].flags & OPF_DATA))
+ continue;
+ if (!is_opr_modified(opr, &ops[i]))
+ continue;
+
+ if (*op_i >= 0) {
+ if (*op_i == i)
+ return 1;
+ // XXX: could check if the other op does the same
+ return -1;
+ }
+
+ *op_i = i;
+ return 1;
+ }
+}
+
+static int try_resolve_const(int i, const struct parsed_opr *opr,
+ int magic, unsigned int *val)
+{
+ int s_i = -1;
+ int ret;
+
+ ret = resolve_origin(i, opr, magic, &s_i, NULL);
+ if (ret == 1) {
+ i = s_i;
+ if (ops[i].op != OP_MOV && ops[i].operand[1].type != OPT_CONST)
+ return -1;
+
+ *val = ops[i].operand[1].val;
+ return 1;
+ }
+
+ return -1;
+}
+
+static struct parsed_proto *process_call_early(int i, int opcnt,
+ int *adj_i)
+{
+ struct parsed_op *po = &ops[i];
+ struct parsed_proto *pp;
+ int multipath = 0;
+ int adj = 0;
+ int ret;
+
+ pp = po->pp;
+ if (pp == NULL || pp->is_vararg || pp->argc_reg != 0)
+ // leave for later
+ return NULL;
+
+ // look for and make use of esp adjust
+ *adj_i = ret = -1;
+ if (!pp->is_stdcall && pp->argc_stack > 0)
+ ret = scan_for_esp_adjust(i + 1, opcnt,
+ pp->argc_stack * 4, &adj, &multipath);
+ if (ret >= 0) {
+ if (pp->argc_stack > adj / 4)
+ return NULL;
+ if (multipath)
+ return NULL;
+ if (ops[ret].op == OP_POP && adj != 4)
+ return NULL;
+ }
+
+ *adj_i = ret;
+ return pp;
+}
+
+static void patch_esp_adjust(struct parsed_op *po, int adj)
+{
+ ferr_assert(po, po->op == OP_ADD);
+ ferr_assert(po, IS(opr_name(po, 0), "esp"));
+ ferr_assert(po, po->operand[1].type == OPT_CONST);
+
+ // this is a bit of a hack, but deals with use of
+ // single adj for multiple calls
+ po->operand[1].val -= adj;
+ po->flags |= OPF_RMD;
+ if (po->operand[1].val == 0)
+ po->flags |= OPF_DONE;
+ ferr_assert(po, (int)po->operand[1].val >= 0);
+}
+
+static struct parsed_proto *process_call(int i, int opcnt)
+{
+ struct parsed_op *po = &ops[i];
+ const struct parsed_proto *pp_c;
+ struct parsed_proto *pp;
+ const char *tmpname;
+ int adj = 0, multipath = 0;
+ int ret, arg;
+
+ tmpname = opr_name(po, 0);
+ pp = po->pp;
+ if (pp == NULL)
+ {
+ // indirect call
+ pp_c = resolve_icall(i, opcnt, &multipath);
+ if (pp_c != NULL) {
+ if (!pp_c->is_func && !pp_c->is_fptr)
+ ferr(po, "call to non-func: %s\n", pp_c->name);
+ pp = proto_clone(pp_c);
+ my_assert_not(pp, NULL);
+ if (multipath)
+ // not resolved just to single func
+ pp->is_fptr = 1;
+
+ switch (po->operand[0].type) {
+ case OPT_REG:
+ // we resolved this call and no longer need the register
+ po->regmask_src &= ~(1 << po->operand[0].reg);
+ break;
+ case OPT_REGMEM:
+ pp->is_fptr = 1;
+ break;
+ default:
+ break;
+ }
+ }
+ if (pp == NULL) {
+ pp = calloc(1, sizeof(*pp));
+ my_assert_not(pp, NULL);
+
+ pp->is_fptr = 1;
+ ret = scan_for_esp_adjust(i + 1, opcnt, 32*4, &adj, &multipath);
+ if (ret < 0 || adj < 0) {
+ if (!g_allow_regfunc)
+ ferr(po, "non-__cdecl indirect call unhandled yet\n");
+ pp->is_unresolved = 1;
+ adj = 0;
+ }
+ adj /= 4;
+ if (adj > ARRAY_SIZE(pp->arg))
+ ferr(po, "esp adjust too large: %d\n", adj);
+ pp->ret_type.name = strdup("int");
+ pp->argc = pp->argc_stack = adj;
+ for (arg = 0; arg < pp->argc; arg++)
+ pp->arg[arg].type.name = strdup("int");
+ }
+ po->pp = pp;
+ }
+
+ // look for and make use of esp adjust
+ multipath = 0;
+ ret = -1;
+ if (!pp->is_stdcall && pp->argc_stack > 0)
+ ret = scan_for_esp_adjust(i + 1, opcnt,
+ pp->argc_stack * 4, &adj, &multipath);
+ if (ret >= 0) {
+ if (pp->is_vararg) {
+ if (adj / 4 < pp->argc_stack) {
+ fnote(po, "(this call)\n");
+ ferr(&ops[ret], "esp adjust is too small: %x < %x\n",
+ adj, pp->argc_stack * 4);
+ }
+ // modify pp to make it have varargs as normal args
+ arg = pp->argc;
+ pp->argc += adj / 4 - pp->argc_stack;
+ for (; arg < pp->argc; arg++) {
+ pp->arg[arg].type.name = strdup("int");
+ pp->argc_stack++;
+ }
+ if (pp->argc > ARRAY_SIZE(pp->arg))
+ ferr(po, "too many args for '%s'\n", tmpname);
+ }
+ if (pp->argc_stack > adj / 4) {
+ fnote(po, "(this call)\n");
+ ferr(&ops[ret], "stack tracking failed for '%s': %x %x\n",
+ tmpname, pp->argc_stack * 4, adj);
+ }
+
+ ops[ret].flags |= OPF_RMD;
+ if (ops[ret].op == OP_POP) {
+ if (adj > 4) {
+ // deal with multi-pop stack adjust
+ adj = pp->argc_stack;
+ while (ops[ret].op == OP_POP && adj > 0 && ret < opcnt) {
+ ops[ret].flags |= OPF_RMD | OPF_DONE;
+ adj--;
+ ret++;
+ }
+ }
+ }
+ else if (!multipath)
+ patch_esp_adjust(&ops[ret], pp->argc_stack * 4);
+ }
+ else if (pp->is_vararg)
+ ferr(po, "missing esp_adjust for vararg func '%s'\n",
+ pp->name);
+
+ return pp;
+}
+
+static int collect_call_args_early(struct parsed_op *po, int i,
+ struct parsed_proto *pp, int *regmask)
+{
+ int arg, ret;
+ int j;
+
+ for (arg = 0; arg < pp->argc; arg++)
+ if (pp->arg[arg].reg == NULL)
+ break;
+
+ // first see if it can be easily done
+ for (j = i; j > 0 && arg < pp->argc; )
+ {
+ if (g_labels[j] != NULL)
+ return -1;
+ j--;
+
+ if (ops[j].op == OP_CALL)
+ return -1;
+ else if (ops[j].op == OP_ADD && ops[j].operand[0].reg == xSP)
+ return -1;
+ else if (ops[j].op == OP_POP)
+ return -1;
+ else if (ops[j].flags & OPF_CJMP)
+ return -1;
+ else if (ops[j].op == OP_PUSH) {
+ if (ops[j].flags & (OPF_FARG|OPF_FARGNR))
+ return -1;
+ ret = scan_for_mod(&ops[j], j + 1, i, 1);
+ if (ret >= 0)
+ return -1;
+
+ if (pp->arg[arg].type.is_va_list)
+ return -1;
+
+ // next arg
+ for (arg++; arg < pp->argc; arg++)
+ if (pp->arg[arg].reg == NULL)
+ break;
+ }
+ }
+
+ if (arg < pp->argc)
+ return -1;
+
+ // now do it
+ for (arg = 0; arg < pp->argc; arg++)
+ if (pp->arg[arg].reg == NULL)
+ break;
+
+ for (j = i; j > 0 && arg < pp->argc; )
+ {
+ j--;
+
+ if (ops[j].op == OP_PUSH)
+ {
+ ops[j].p_argnext = -1;
+ ferr_assert(&ops[j], pp->arg[arg].datap == NULL);
+ pp->arg[arg].datap = &ops[j];
+
+ if (ops[j].operand[0].type == OPT_REG)
+ *regmask |= 1 << ops[j].operand[0].reg;
+
+ ops[j].flags |= OPF_RMD | OPF_DONE | OPF_FARGNR | OPF_FARG;
+ ops[j].flags &= ~OPF_RSAVE;
+
+ // next arg
+ for (arg++; arg < pp->argc; arg++)
+ if (pp->arg[arg].reg == NULL)
+ break;
+ }
+ }
+
+ return 0;
+}
+
+static int collect_call_args_r(struct parsed_op *po, int i,
+ struct parsed_proto *pp, int *regmask, int *save_arg_vars,
+ int *arg_grp, int arg, int magic, int need_op_saving, int may_reuse)
+{
+ struct parsed_proto *pp_tmp;
+ struct parsed_op *po_tmp;
+ struct label_ref *lr;
+ int need_to_save_current;
+ int arg_grp_current = 0;
+ int save_args_seen = 0;
+ int save_args;
+ int ret = 0;
+ int reg;
+ char buf[32];
+ int j, k;
+
+ if (i < 0) {
+ ferr(po, "dead label encountered\n");
+ return -1;
+ }
+
+ for (; arg < pp->argc; arg++)
+ if (pp->arg[arg].reg == NULL)
+ break;
+ magic = (magic & 0xffffff) | (arg << 24);
+
+ for (j = i; j >= 0 && (arg < pp->argc || pp->is_unresolved); )
+ {
+ if (((ops[j].cc_scratch ^ magic) & 0xffffff) == 0) {
+ if (ops[j].cc_scratch != magic) {
+ ferr(&ops[j], "arg collect hit same path with diff args for %s\n",
+ pp->name);
+ return -1;
+ }
+ // ok: have already been here
+ return 0;
+ }
+ ops[j].cc_scratch = magic;
+
+ if (g_labels[j] != NULL && g_label_refs[j].i != -1) {
+ lr = &g_label_refs[j];
+ if (lr->next != NULL)
+ need_op_saving = 1;
+ for (; lr->next; lr = lr->next) {
+ check_i(&ops[j], lr->i);
+ 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_grp, arg, magic, need_op_saving, may_reuse);
+ if (ret < 0)
+ return ret;
+ }
+
+ check_i(&ops[j], lr->i);
+ 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
+ j = lr->i;
+ continue;
+ }
+ need_op_saving = 1;
+ ret = collect_call_args_r(po, lr->i, pp, regmask, save_arg_vars,
+ arg_grp, arg, magic, need_op_saving, may_reuse);
+ if (ret < 0)
+ return ret;
+ }
+ j--;
+
+ if (ops[j].op == OP_CALL)
+ {
+ if (pp->is_unresolved)
+ break;
+
+ pp_tmp = ops[j].pp;
+ if (pp_tmp == NULL)
+ ferr(po, "arg collect hit unparsed call '%s'\n",
+ ops[j].operand[0].name);
+ if (may_reuse && pp_tmp->argc_stack > 0)
+ ferr(po, "arg collect %d/%d hit '%s' with %d stack args\n",
+ arg, pp->argc, opr_name(&ops[j], 0), pp_tmp->argc_stack);
+ }
+ // esp adjust of 0 means we collected it before
+ else if (ops[j].op == OP_ADD && ops[j].operand[0].reg == xSP
+ && (ops[j].operand[1].type != OPT_CONST
+ || ops[j].operand[1].val != 0))
+ {
+ if (pp->is_unresolved)
+ break;
+
+ ferr(po, "arg collect %d/%d hit esp adjust of %d\n",
+ arg, pp->argc, ops[j].operand[1].val);
+ }
+ else if (ops[j].op == OP_POP && !(ops[j].flags & OPF_DONE))
+ {
+ if (pp->is_unresolved)
+ break;
+
+ ferr(po, "arg collect %d/%d hit pop\n", arg, pp->argc);
+ }
+ else if (ops[j].flags & OPF_CJMP)
+ {
+ if (pp->is_unresolved)
+ break;
+
+ may_reuse = 1;
+ }
+ else if (ops[j].op == OP_PUSH
+ && !(ops[j].flags & (OPF_FARGNR|OPF_DONE)))
+ {
+ if (pp->is_unresolved && (ops[j].flags & OPF_RMD))
+ break;
+
+ ops[j].p_argnext = -1;
+ po_tmp = pp->arg[arg].datap;
+ if (po_tmp != NULL)
+ ops[j].p_argnext = po_tmp - ops;
+ pp->arg[arg].datap = &ops[j];
+
+ need_to_save_current = 0;
+ save_args = 0;
+ reg = -1;
+ if (ops[j].operand[0].type == OPT_REG)
+ reg = ops[j].operand[0].reg;
+
+ if (!need_op_saving) {
+ ret = scan_for_mod(&ops[j], j + 1, i, 1);
+ need_to_save_current = (ret >= 0);
+ }
+ if (need_op_saving || need_to_save_current) {
+ // mark this push as one that needs operand saving
+ ops[j].flags &= ~OPF_RMD;
+ if (ops[j].p_argnum == 0) {
+ ops[j].p_argnum = arg + 1;
+ save_args |= 1 << arg;
+ }
+ else if (ops[j].p_argnum < arg + 1) {
+ // XXX: might kill valid var..
+ //*save_arg_vars &= ~(1 << (ops[j].p_argnum - 1));
+ ops[j].p_argnum = arg + 1;
+ save_args |= 1 << arg;
+ }
+
+ if (save_args_seen & (1 << (ops[j].p_argnum - 1))) {
+ save_args_seen = 0;
+ arg_grp_current++;
+ if (arg_grp_current >= MAX_ARG_GRP)
+ ferr(&ops[j], "out of arg groups (arg%d), f %s\n",
+ ops[j].p_argnum, pp->name);
+ }
+ }
+ else if (ops[j].p_argnum == 0)
+ ops[j].flags |= OPF_RMD;
+
+ // some PUSHes are reused by different calls on other branches,
+ // but that can't happen if we didn't branch, so they
+ // can be removed from future searches (handles nested calls)
+ if (!may_reuse)
+ ops[j].flags |= OPF_FARGNR;
+
+ ops[j].flags |= OPF_FARG;
+ ops[j].flags &= ~OPF_RSAVE;
+
+ // check for __VALIST
+ if (!pp->is_unresolved && pp->arg[arg].type.is_va_list) {
+ k = -1;
+ ret = resolve_origin(j, &ops[j].operand[0],
+ magic + 1, &k, NULL);
+ if (ret == 1 && k >= 0)
+ {
+ if (ops[k].op == OP_LEA) {
+ snprintf(buf, sizeof(buf), "arg_%X",
+ g_func_pp->argc_stack * 4);
+ if (!g_func_pp->is_vararg
+ || strstr(ops[k].operand[1].name, buf))
+ {
+ ops[k].flags |= OPF_RMD | OPF_DONE;
+ ops[j].flags |= OPF_RMD | OPF_VAPUSH;
+ save_args &= ~(1 << arg);
+ reg = -1;
+ }
+ else
+ ferr(&ops[j], "lea va_list used, but no vararg?\n");
+ }
+ // check for va_list from g_func_pp arg too
+ else if (ops[k].op == OP_MOV
+ && is_stack_access(&ops[k], &ops[k].operand[1]))
+ {
+ ret = stack_frame_access(&ops[k], &ops[k].operand[1],
+ buf, sizeof(buf), ops[k].operand[1].name, "", 1, 0);
+ if (ret >= 0) {
+ ops[k].flags |= OPF_RMD | OPF_DONE;
+ ops[j].flags |= OPF_RMD;
+ ops[j].p_argpass = ret + 1;
+ save_args &= ~(1 << arg);
+ reg = -1;
+ }
+ }
+ }
+ }
+
+ *save_arg_vars |= save_args;
+
+ // tracking reg usage
+ if (reg >= 0)
+ *regmask |= 1 << reg;
+
+ arg++;
+ if (!pp->is_unresolved) {
+ // next arg
+ for (; arg < pp->argc; arg++)
+ if (pp->arg[arg].reg == NULL)
+ break;
+ }
+ magic = (magic & 0xffffff) | (arg << 24);
+ }
+
+ if (ops[j].p_arggrp > arg_grp_current) {
+ save_args_seen = 0;
+ arg_grp_current = ops[j].p_arggrp;
+ }
+ if (ops[j].p_argnum > 0)
+ save_args_seen |= 1 << (ops[j].p_argnum - 1);
+ }
+
+ if (arg < pp->argc) {
+ ferr(po, "arg collect failed for '%s': %d/%d\n",
+ pp->name, arg, pp->argc);
+ return -1;
+ }
+
+ if (arg_grp_current > *arg_grp)
+ *arg_grp = arg_grp_current;
+
+ return arg;
+}
+
+static int collect_call_args(struct parsed_op *po, int i,
+ struct parsed_proto *pp, int *regmask, int *save_arg_vars,
+ int magic)
+{
+ // arg group is for cases when pushes for
+ // multiple funcs are going on
+ struct parsed_op *po_tmp;
+ int save_arg_vars_current = 0;
+ int arg_grp = 0;
+ int ret;
+ int a;
+
+ ret = collect_call_args_r(po, i, pp, regmask,
+ &save_arg_vars_current, &arg_grp, 0, magic, 0, 0);
+ if (ret < 0)
+ return ret;
+
+ if (arg_grp != 0) {
+ // propagate arg_grp
+ for (a = 0; a < pp->argc; a++) {
+ if (pp->arg[a].reg != NULL)
+ continue;
+
+ po_tmp = pp->arg[a].datap;
+ while (po_tmp != NULL) {
+ po_tmp->p_arggrp = arg_grp;
+ if (po_tmp->p_argnext > 0)
+ po_tmp = &ops[po_tmp->p_argnext];
+ else
+ po_tmp = NULL;
+ }
+ }
+ }
+ save_arg_vars[arg_grp] |= save_arg_vars_current;
+
+ if (pp->is_unresolved) {
+ pp->argc += ret;
+ pp->argc_stack += ret;
+ for (a = 0; a < pp->argc; a++)
+ if (pp->arg[a].type.name == NULL)
+ pp->arg[a].type.name = strdup("int");
+ }
+
+ return ret;
+}
+
+static void pp_insert_reg_arg(struct parsed_proto *pp, const char *reg)
+{
+ int i;
+
+ for (i = 0; i < pp->argc; i++)
+ if (pp->arg[i].reg == NULL)
+ break;
+
+ if (pp->argc_stack)
+ memmove(&pp->arg[i + 1], &pp->arg[i],
+ sizeof(pp->arg[0]) * pp->argc_stack);
+ memset(&pp->arg[i], 0, sizeof(pp->arg[i]));
+ pp->arg[i].reg = strdup(reg);
+ pp->arg[i].type.name = strdup("int");
+ pp->argc++;
+ pp->argc_reg++;
+}
+
+static void add_label_ref(struct label_ref *lr, int op_i)
+{
+ struct label_ref *lr_new;
+
+ if (lr->i == -1) {
+ lr->i = op_i;
+ return;
+ }
+
+ lr_new = calloc(1, sizeof(*lr_new));
+ lr_new->i = op_i;
+ lr_new->next = lr->next;
+ lr->next = lr_new;
+}
+
+static struct parsed_data *try_resolve_jumptab(int i, int opcnt)
+{
+ struct parsed_op *po = &ops[i];
+ struct parsed_data *pd;
+ char label[NAMELEN], *p;
+ int len, j, l;
+
+ p = strchr(po->operand[0].name, '[');
+ if (p == NULL)
+ return NULL;
+
+ len = p - po->operand[0].name;
+ strncpy(label, po->operand[0].name, len);
+ label[len] = 0;
+
+ for (j = 0, pd = NULL; j < g_func_pd_cnt; j++) {
+ if (IS(g_func_pd[j].label, label)) {
+ pd = &g_func_pd[j];
+ break;
+ }
+ }
+ if (pd == NULL)
+ //ferr(po, "label '%s' not parsed?\n", label);
+ return NULL;
+
+ if (pd->type != OPT_OFFSET)
+ ferr(po, "label '%s' with non-offset data?\n", label);
+
+ // find all labels, link
+ for (j = 0; j < pd->count; j++) {
+ for (l = 0; l < opcnt; l++) {
+ if (g_labels[l] != NULL && IS(g_labels[l], pd->d[j].u.label)) {
+ add_label_ref(&g_label_refs[l], i);
+ pd->d[j].bt_i = l;
+ break;
+ }
+ }
+ }
+
+ return pd;
+}
+
+static void clear_labels(int count)
+{
+ int i;
+
+ for (i = 0; i < count; i++) {
+ if (g_labels[i] != NULL) {
+ free(g_labels[i]);
+ g_labels[i] = NULL;
+ }
+ }
+}
+
+static void output_std_flags(FILE *fout, struct parsed_op *po,
+ int *pfomask, const char *dst_opr_text)
+{
+ if (*pfomask & (1 << PFO_Z)) {
+ fprintf(fout, "\n cond_z = (%s%s == 0);",
+ lmod_cast_u(po, po->operand[0].lmod), dst_opr_text);
+ *pfomask &= ~(1 << PFO_Z);
+ }
+ if (*pfomask & (1 << PFO_S)) {
+ fprintf(fout, "\n cond_s = (%s%s < 0);",
+ lmod_cast_s(po, po->operand[0].lmod), dst_opr_text);
+ *pfomask &= ~(1 << PFO_S);
+ }
+}
+
+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 int get_pp_arg_regmask(const struct parsed_proto *pp)
+{
+ int regmask = 0;
+ int i, reg;
+
+ for (i = 0; i < pp->argc; i++) {
+ if (pp->arg[i].reg != NULL) {
+ reg = char_array_i(regs_r32,
+ ARRAY_SIZE(regs_r32), pp->arg[i].reg);
+ if (reg < 0)
+ ferr(ops, "arg '%s' of func '%s' is not a reg?\n",
+ pp->arg[i].reg, pp->name);
+ regmask |= 1 << reg;
+ }
+ }
+
+ return regmask;
+}
+
+static char *saved_arg_name(char *buf, size_t buf_size, int grp, int num)
+{
+ char buf1[16];
+
+ buf1[0] = 0;
+ if (grp > 0)
+ snprintf(buf1, sizeof(buf1), "%d", grp);
+ snprintf(buf, buf_size, "s%s_a%d", buf1, num);
+
+ return buf;
+}
+
+static void gen_x_cleanup(int opcnt);
+
+static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt)
+{
+ struct parsed_op *po, *delayed_flag_op = NULL, *tmp_op;
+ struct parsed_opr *last_arith_dst = NULL;
+ char buf1[256], buf2[256], buf3[256], cast[64];
+ const struct parsed_proto *pp_c;
+ struct parsed_proto *pp, *pp_tmp;
+ struct parsed_data *pd;
+ const char *tmpname;
+ unsigned int uval;
+ int save_arg_vars[MAX_ARG_GRP] = { 0, };
+ int cond_vars = 0;
+ int need_tmp_var = 0;
+ int need_tmp64 = 0;
+ int had_decl = 0;
+ int label_pending = 0;
+ int regmask_save = 0;
+ int regmask_arg = 0;
+ int regmask_now = 0;
+ int regmask_init = 0;
+ int regmask = 0;
+ int pfomask = 0;
+ int found = 0;
+ int depth = 0;
+ int no_output;
+ int i, j, l;
+ int arg;
+ int reg;
+ int ret;
+
+ g_bp_frame = g_sp_frame = g_stack_fsz = 0;
+ g_stack_frame_used = 0;
+
+ g_func_pp = proto_parse(fhdr, funcn, 0);
+ if (g_func_pp == NULL)
+ ferr(ops, "proto_parse failed for '%s'\n", funcn);
+
+ regmask_arg = get_pp_arg_regmask(g_func_pp);
+
+ // pass1:
+ // - handle ebp/esp frame, remove ops related to it
+ scan_prologue_epilogue(opcnt);
+
+ // pass2:
+ // - parse calls with labels
+ // - resolve all branches
+ for (i = 0; i < opcnt; i++)
+ {
+ po = &ops[i];
+ po->bt_i = -1;
+ po->btj = NULL;
+
+ if (po->flags & (OPF_RMD|OPF_DONE))
+ continue;
+
+ if (po->op == OP_CALL) {
+ pp = NULL;
+
+ if (po->operand[0].type == OPT_LABEL) {
+ tmpname = opr_name(po, 0);
+ if (IS_START(tmpname, "loc_"))
+ ferr(po, "call to loc_*\n");
+ pp_c = proto_parse(fhdr, tmpname, 0);
+ if (pp_c == NULL)
+ ferr(po, "proto_parse failed for call '%s'\n", tmpname);
+
+ pp = proto_clone(pp_c);
+ my_assert_not(pp, NULL);
+ }
+ else if (po->datap != NULL) {
+ pp = calloc(1, sizeof(*pp));
+ my_assert_not(pp, NULL);
+
+ ret = parse_protostr(po->datap, pp);
+ if (ret < 0)
+ ferr(po, "bad protostr supplied: %s\n", (char *)po->datap);
+ free(po->datap);
+ po->datap = NULL;
+ }
+
+ if (pp != NULL) {
+ if (pp->is_fptr)
+ check_func_pp(po, pp, "fptr var call");
+ if (pp->is_noreturn)
+ po->flags |= OPF_TAIL;
+ }
+ po->pp = pp;
+ continue;
+ }
+
+ if (!(po->flags & OPF_JMP) || po->op == OP_RET)
+ continue;
+
+ if (po->operand[0].type == OPT_REGMEM) {
+ pd = try_resolve_jumptab(i, opcnt);
+ if (pd == NULL)
+ goto tailcall;
+
+ po->btj = pd;
+ continue;
+ }
+
+ for (l = 0; l < opcnt; l++) {
+ if (g_labels[l] != NULL
+ && IS(po->operand[0].name, g_labels[l]))
+ {
+ if (l == i + 1 && po->op == OP_JMP) {
+ // yet another alignment type..
+ po->flags |= OPF_RMD|OPF_DONE;
+ break;
+ }
+ add_label_ref(&g_label_refs[l], i);
+ po->bt_i = l;
+ break;
+ }
+ }
+
+ if (po->bt_i != -1 || (po->flags & OPF_RMD))
+ continue;
+
+ if (po->operand[0].type == OPT_LABEL)
+ // assume tail call
+ goto tailcall;
+
+ ferr(po, "unhandled branch\n");
+
+tailcall:
+ po->op = OP_CALL;
+ po->flags |= OPF_TAIL;
+ if (i > 0 && ops[i - 1].op == OP_POP)
+ po->flags |= OPF_ATAIL;
+ i--; // reprocess
+ }
+
+ // pass3:
+ // - remove dead labels
+ // - process trivial calls
+ for (i = 0; i < opcnt; i++)
+ {
+ if (g_labels[i] != NULL && g_label_refs[i].i == -1) {
+ free(g_labels[i]);
+ g_labels[i] = NULL;
+ }
+
+ po = &ops[i];
+ if (po->flags & (OPF_RMD|OPF_DONE))
+ continue;
+
+ if (po->op == OP_CALL)
+ {
+ pp = process_call_early(i, opcnt, &j);
+ if (pp != NULL) {
+ if (!(po->flags & OPF_ATAIL))
+ // since we know the args, try to collect them
+ if (collect_call_args_early(po, i, pp, ®mask) != 0)
+ pp = NULL;
+ }
+
+ if (pp != NULL) {
+ if (j >= 0) {
+ // commit esp adjust
+ ops[j].flags |= OPF_RMD;
+ if (ops[j].op != OP_POP)
+ patch_esp_adjust(&ops[j], pp->argc_stack * 4);
+ else
+ ops[j].flags |= OPF_DONE;
+ }
+
+ if (strstr(pp->ret_type.name, "int64"))
+ need_tmp64 = 1;
+
+ po->flags |= OPF_DONE;
+ }
+ }
+ }
+
+ // pass4:
+ // - process calls
+ for (i = 0; i < opcnt; i++)
+ {
+ po = &ops[i];
+ if (po->flags & (OPF_RMD|OPF_DONE))
+ continue;
+
+ if (po->op == OP_CALL && !(po->flags & OPF_DONE))
+ {
+ pp = process_call(i, opcnt);
+
+ if (!pp->is_unresolved && !(po->flags & OPF_ATAIL)) {
+ // since we know the args, collect them
+ collect_call_args(po, i, pp, ®mask, save_arg_vars,
+ i + opcnt * 2);
+ }
+
+ if (strstr(pp->ret_type.name, "int64"))
+ need_tmp64 = 1;
+ }
+ }
+
+ // pass5:
+ // - 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
+ // - declare indirect functions
+ for (i = 0; i < opcnt; i++)
+ {
+ po = &ops[i];
+ if (po->flags & (OPF_RMD|OPF_DONE))
+ continue;
+
+ if (po->op == OP_PUSH && (po->flags & OPF_RSAVE)) {
+ reg = po->operand[0].reg;
+ if (!(regmask & (1 << reg)))
+ // not a reg save after all, rerun scan_for_pop
+ po->flags &= ~OPF_RSAVE;
+ else
+ regmask_save |= 1 << reg;
+ }
+
+ if (po->op == OP_PUSH && !(po->flags & OPF_FARG)
+ && !(po->flags & OPF_RSAVE) && !g_func_pp->is_userstack)
+ {
+ 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;
+ }
+ }
+ else if (po->operand[0].type == OPT_CONST) {
+ scan_for_pop_const(i, opcnt);
+ }
+ }
+
+ if (po->op == OP_STD) {
+ po->flags |= OPF_DF | OPF_RMD | OPF_DONE;
+ 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)) {
+ if (po->regmask_dst & (1 << xBP))
+ // compiler decided to drop bp frame and use ebp as scratch
+ scan_fwd_set_flags(i + 1, opcnt, i + opcnt * 5, OPF_EBP_S);
+ else
+ regmask_now &= ~(1 << xBP);
+ }
+ }
+
+ regmask |= regmask_now;
+
+ if (po->flags & OPF_CC)
+ {
+ int setters[16], cnt = 0, branched = 0;
+
+ ret = scan_for_flag_set(i, i + opcnt * 6,
+ &branched, setters, &cnt);
+ if (ret < 0 || cnt <= 0)
+ ferr(po, "unable to trace flag setter(s)\n");
+ if (cnt > ARRAY_SIZE(setters))
+ ferr(po, "too many flag setters\n");
+
+ for (j = 0; j < cnt; j++)
+ {
+ tmp_op = &ops[setters[j]]; // flag setter
+ pfomask = 0;
+
+ // to get nicer code, we try to delay test and cmp;
+ // if we can't because of operand modification, or if we
+ // 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 << po->pfo;
+ }
+ else if (tmp_op->op == OP_CMPS || tmp_op->op == OP_SCAS) {
+ pfomask = 1 << po->pfo;
+ }
+ else {
+ // see if we'll be able to handle based on op result
+ if ((tmp_op->op != OP_AND && tmp_op->op != OP_OR
+ && po->pfo != PFO_Z && po->pfo != PFO_S
+ && po->pfo != PFO_P)
+ || branched
+ || scan_for_mod_opr0(tmp_op, setters[j] + 1, i) >= 0)
+ {
+ pfomask = 1 << po->pfo;
+ }
+
+ if (tmp_op->op == OP_ADD && po->pfo == PFO_C) {
+ propagate_lmod(tmp_op, &tmp_op->operand[0],
+ &tmp_op->operand[1]);
+ if (tmp_op->operand[0].lmod == OPLM_DWORD)
+ need_tmp64 = 1;
+ }
+ }
+ if (pfomask) {
+ tmp_op->pfomask |= pfomask;
+ cond_vars |= pfomask;
+ }
+ // note: may overwrite, currently not a problem
+ po->datap = tmp_op;
+ }
+
+ if (po->op == OP_RCL || po->op == OP_RCR
+ || po->op == OP_ADC || po->op == OP_SBB)
+ cond_vars |= 1 << PFO_C;
+ }
+
+ if (po->op == OP_CMPS || po->op == OP_SCAS) {
+ cond_vars |= 1 << PFO_Z;
+ }
+ else if (po->op == OP_MUL
+ || (po->op == OP_IMUL && po->operand_cnt == 1))
+ {
+ if (po->operand[0].lmod == OPLM_DWORD)
+ need_tmp64 = 1;
+ }
+ else if (po->op == OP_CALL) {
+ // note: resolved non-reg calls are OPF_DONE already
+ pp = po->pp;
+ if (pp == NULL)
+ ferr(po, "NULL pp\n");
+
+ if (pp->is_unresolved) {
+ int regmask_stack = 0;
+ 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->p_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)|(1 << xDX))))
+ {
+ pp_insert_reg_arg(pp, "ecx");
+ pp->is_fastcall = 1;
+ regmask_init |= 1 << xCX;
+ regmask |= 1 << xCX;
+ }
+ if (pp->argc_stack != 0
+ || ((regmask | regmask_arg) & (1 << xDX)))
+ {
+ pp_insert_reg_arg(pp, "edx");
+ regmask_init |= 1 << xDX;
+ regmask |= 1 << xDX;
+ }
+ }
+
+ // 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++) {
+ if (pp->arg[arg].reg != NULL) {
+ reg = char_array_i(regs_r32,
+ ARRAY_SIZE(regs_r32), pp->arg[arg].reg);
+ if (reg < 0)
+ ferr(ops, "arg '%s' is not a reg?\n", pp->arg[arg].reg);
+ if (!(regmask & (1 << reg))) {
+ regmask_init |= 1 << reg;
+ regmask |= 1 << reg;
+ }
+ }
+ }
+ }
+ else if (po->op == OP_MOV && po->operand[0].pp != NULL
+ && po->operand[1].pp != NULL)
+ {
+ // <var> = offset <something>
+ if ((po->operand[1].pp->is_func || po->operand[1].pp->is_fptr)
+ && !IS_START(po->operand[1].name, "off_"))
+ {
+ if (!po->operand[0].pp->is_fptr)
+ ferr(po, "%s not declared as fptr when it should be\n",
+ po->operand[0].name);
+ if (pp_cmp_func(po->operand[0].pp, po->operand[1].pp)) {
+ pp_print(buf1, sizeof(buf1), po->operand[0].pp);
+ pp_print(buf2, sizeof(buf2), po->operand[1].pp);
+ fnote(po, "var: %s\n", buf1);
+ fnote(po, "func: %s\n", buf2);
+ ferr(po, "^ mismatch\n");
+ }
+ }
+ }
+ else if (po->op == OP_RET && !IS(g_func_pp->ret_type.name, "void"))
+ regmask |= 1 << xAX;
+ else if (po->op == OP_DIV || po->op == OP_IDIV) {
+ // 32bit division is common, look for it
+ if (po->op == OP_DIV)
+ ret = scan_for_reg_clear(i, xDX);
+ else
+ ret = scan_for_cdq_edx(i);
+ if (ret >= 0)
+ po->flags |= OPF_32BIT;
+ else
+ need_tmp64 = 1;
+ }
+ else if (po->op == OP_CLD)
+ po->flags |= OPF_RMD | OPF_DONE;
+
+ if (po->op == OP_RCL || po->op == OP_RCR || po->op == OP_XCHG) {
+ need_tmp_var = 1;
+ }
+ }
+
+ // pass6:
+ // - confirm regmask_save, it might have been reduced
+ if (regmask_save != 0)
+ {
+ regmask_save = 0;
+ for (i = 0; i < opcnt; i++) {
+ po = &ops[i];
+ if (po->flags & OPF_RMD)
+ continue;
+
+ if (po->op == OP_PUSH && (po->flags & OPF_RSAVE))
+ regmask_save |= 1 << po->operand[0].reg;
+ }
+ }
+
+ // output starts here
+
+ // define userstack size
+ if (g_func_pp->is_userstack) {
+ fprintf(fout, "#ifndef US_SZ_%s\n", g_func_pp->name);
+ fprintf(fout, "#define US_SZ_%s USERSTACK_SIZE\n", g_func_pp->name);
+ fprintf(fout, "#endif\n");
+ }
+
+ // the function itself
+ fprintf(fout, "%s ", g_func_pp->ret_type.name);
+ output_pp_attrs(fout, g_func_pp, g_ida_func_attr & IDAFA_NORETURN);
+ fprintf(fout, "%s(", g_func_pp->name);
+
+ for (i = 0; i < g_func_pp->argc; i++) {
+ if (i > 0)
+ fprintf(fout, ", ");
+ if (g_func_pp->arg[i].fptr != NULL) {
+ // func pointer..
+ pp = g_func_pp->arg[i].fptr;
+ fprintf(fout, "%s (", pp->ret_type.name);
+ output_pp_attrs(fout, pp, 0);
+ fprintf(fout, "*a%d)(", i + 1);
+ for (j = 0; j < pp->argc; j++) {
+ if (j > 0)
+ fprintf(fout, ", ");
+ if (pp->arg[j].fptr)
+ ferr(ops, "nested fptr\n");
+ fprintf(fout, "%s", pp->arg[j].type.name);
+ }
+ if (pp->is_vararg) {
+ if (j > 0)
+ fprintf(fout, ", ");
+ fprintf(fout, "...");
+ }
+ fprintf(fout, ")");
+ }
+ else if (g_func_pp->arg[i].type.is_retreg) {
+ fprintf(fout, "u32 *r_%s", g_func_pp->arg[i].reg);
+ }
+ else {
+ fprintf(fout, "%s a%d", g_func_pp->arg[i].type.name, i + 1);
+ }
+ }
+ if (g_func_pp->is_vararg) {
+ if (i > 0)
+ fprintf(fout, ", ");
+ fprintf(fout, "...");
+ }
+
+ fprintf(fout, ")\n{\n");
+
+ // declare indirect functions
+ for (i = 0; i < opcnt; i++) {
+ po = &ops[i];
+ if (po->flags & OPF_RMD)
+ continue;
+
+ if (po->op == OP_CALL) {
+ pp = po->pp;
+ if (pp == NULL)
+ ferr(po, "NULL pp\n");
+
+ if (pp->is_fptr && !(pp->name[0] != 0 && pp->is_arg)) {
+ if (pp->name[0] != 0) {
+ memmove(pp->name + 2, pp->name, strlen(pp->name) + 1);
+ memcpy(pp->name, "i_", 2);
+
+ // might be declared already
+ found = 0;
+ for (j = 0; j < i; j++) {
+ if (ops[j].op == OP_CALL && (pp_tmp = ops[j].pp)) {
+ if (pp_tmp->is_fptr && IS(pp->name, pp_tmp->name)) {
+ found = 1;
+ break;
+ }
+ }
+ }
+ if (found)
+ continue;
+ }
+ else
+ snprintf(pp->name, sizeof(pp->name), "icall%d", i);
+
+ fprintf(fout, " %s (", pp->ret_type.name);
+ output_pp_attrs(fout, pp, 0);
+ fprintf(fout, "*%s)(", pp->name);
+ for (j = 0; j < pp->argc; j++) {
+ if (j > 0)
+ fprintf(fout, ", ");
+ fprintf(fout, "%s a%d", pp->arg[j].type.name, j + 1);
+ }
+ fprintf(fout, ");\n");
+ }
+ }
+ }
+
+ // output LUTs/jumptables
+ for (i = 0; i < g_func_pd_cnt; i++) {
+ pd = &g_func_pd[i];
+ fprintf(fout, " static const ");
+ if (pd->type == OPT_OFFSET) {
+ fprintf(fout, "void *jt_%s[] =\n { ", pd->label);
+
+ for (j = 0; j < pd->count; j++) {
+ if (j > 0)
+ fprintf(fout, ", ");
+ fprintf(fout, "&&%s", pd->d[j].u.label);
+ }
+ }
+ else {
+ fprintf(fout, "%s %s[] =\n { ",
+ lmod_type_u(ops, pd->lmod), pd->label);
+
+ for (j = 0; j < pd->count; j++) {
+ if (j > 0)
+ fprintf(fout, ", ");
+ fprintf(fout, "%u", pd->d[j].u.val);
+ }
+ }
+ fprintf(fout, " };\n");
+ had_decl = 1;
+ }
+
+ // declare stack frame, va_arg
+ if (g_stack_fsz) {
+ fprintf(fout, " union { u32 d[%d]; u16 w[%d]; u8 b[%d]; } sf;\n",
+ (g_stack_fsz + 3) / 4, (g_stack_fsz + 1) / 2, g_stack_fsz);
+ had_decl = 1;
+ }
+
+ if (g_func_pp->is_userstack) {
+ fprintf(fout, " u32 fake_sf[US_SZ_%s / 4];\n", g_func_pp->name);
+ fprintf(fout, " u32 *esp = &fake_sf[sizeof(fake_sf) / 4];\n");
+ had_decl = 1;
+ }
+
+ if (g_func_pp->is_vararg) {
+ fprintf(fout, " va_list ap;\n");
+ had_decl = 1;
+ }
+
+ // declare arg-registers
+ for (i = 0; i < g_func_pp->argc; i++) {
+ if (g_func_pp->arg[i].reg != NULL) {
+ reg = char_array_i(regs_r32,
+ ARRAY_SIZE(regs_r32), g_func_pp->arg[i].reg);
+ if (regmask & (1 << reg)) {
+ if (g_func_pp->arg[i].type.is_retreg)
+ fprintf(fout, " u32 %s = *r_%s;\n",
+ g_func_pp->arg[i].reg, g_func_pp->arg[i].reg);
+ else
+ fprintf(fout, " u32 %s = (u32)a%d;\n",
+ g_func_pp->arg[i].reg, i + 1);
+ }
+ else {
+ if (g_func_pp->arg[i].type.is_retreg)
+ ferr(ops, "retreg '%s' is unused?\n",
+ g_func_pp->arg[i].reg);
+ fprintf(fout, " // %s = a%d; // unused\n",
+ g_func_pp->arg[i].reg, i + 1);
+ }
+ had_decl = 1;
+ }
+ }
+
+ regmask_now = regmask & ~regmask_arg;
+ regmask_now &= ~(1 << xSP);
+ if (regmask_now & 0x00ff) {
+ for (reg = 0; reg < 8; reg++) {
+ if (regmask_now & (1 << reg)) {
+ fprintf(fout, " u32 %s", regs_r32[reg]);
+ if (regmask_init & (1 << reg))
+ fprintf(fout, " = 0");
+ fprintf(fout, ";\n");
+ had_decl = 1;
+ }
+ }
+ }
+ if (regmask_now & 0xff00) {
+ for (reg = 8; reg < 16; reg++) {
+ if (regmask_now & (1 << reg)) {
+ fprintf(fout, " mmxr %s", regs_r32[reg]);
+ if (regmask_init & (1 << reg))
+ fprintf(fout, " = { 0, }");
+ fprintf(fout, ";\n");
+ had_decl = 1;
+ }
+ }
+ }
+
+ if (regmask_save) {
+ for (reg = 0; reg < 8; reg++) {
+ if (regmask_save & (1 << reg)) {
+ fprintf(fout, " u32 s_%s;\n", regs_r32[reg]);
+ had_decl = 1;
+ }
+ }
+ }
+
+ for (i = 0; i < ARRAY_SIZE(save_arg_vars); i++) {
+ if (save_arg_vars[i] == 0)
+ continue;
+ for (reg = 0; reg < 32; reg++) {
+ if (save_arg_vars[i] & (1 << reg)) {
+ fprintf(fout, " u32 %s;\n",
+ saved_arg_name(buf1, sizeof(buf1), i, reg + 1));
+ had_decl = 1;
+ }
+ }
+ }
+
+ if (cond_vars) {
+ for (i = 0; i < 8; i++) {
+ if (cond_vars & (1 << i)) {
+ fprintf(fout, " u32 cond_%s;\n", parsed_flag_op_names[i]);
+ had_decl = 1;
+ }
+ }
+ }
+
+ if (need_tmp_var) {
+ fprintf(fout, " u32 tmp;\n");
+ had_decl = 1;
+ }
+
+ if (need_tmp64) {
+ fprintf(fout, " u64 tmp64;\n");
+ had_decl = 1;
+ }
+
+ if (had_decl)
+ fprintf(fout, "\n");
+
+ if (g_func_pp->is_vararg) {
+ if (g_func_pp->argc_stack == 0)
+ ferr(ops, "vararg func without stack args?\n");
+ fprintf(fout, " va_start(ap, a%d);\n", g_func_pp->argc);
+ }
+
+ // output ops
+ for (i = 0; i < opcnt; i++)
+ {
+ if (g_labels[i] != NULL) {
+ fprintf(fout, "\n%s:\n", g_labels[i]);
+ label_pending = 1;
+
+ delayed_flag_op = NULL;
+ last_arith_dst = NULL;
+ }
+
+ po = &ops[i];
+ if (po->flags & OPF_RMD)
+ continue;
+
+ no_output = 0;
+
+ #define assert_operand_cnt(n_) \
+ if (po->operand_cnt != n_) \
+ ferr(po, "operand_cnt is %d/%d\n", po->operand_cnt, n_)
+
+ // conditional/flag using op?
+ if (po->flags & OPF_CC)
+ {
+ int is_delayed = 0;
+
+ tmp_op = po->datap;
+
+ // we go through all this trouble to avoid using parsed_flag_op,
+ // which makes generated code much nicer
+ if (delayed_flag_op != NULL)
+ {
+ out_cmp_test(buf1, sizeof(buf1), delayed_flag_op,
+ po->pfo, po->pfo_inv);
+ is_delayed = 1;