+ po->btj = pd;
+ continue;
+ }
+
+ for (l = 0; l < opcnt; l++) {
+ if (g_labels[l][0] && IS(po->operand[0].name, g_labels[l])) {
+ if (l == i + 1 && po->op == OP_JMP) {
+ // yet another alignment type..
+ po->flags |= OPF_RMD;
+ 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 calls
+ for (i = 0; i < opcnt; i++)
+ {
+ if (g_labels[i][0] != 0 && g_label_refs[i].i == -1)
+ g_labels[i][0] = 0;
+
+ po = &ops[i];
+ if (po->flags & OPF_RMD)
+ continue;
+
+ if (po->op == OP_CALL)
+ {
+ tmpname = opr_name(po, 0);
+ pp = po->pp;
+ if (pp == NULL)
+ {
+ // indirect call
+ pp_c = resolve_icall(i, opcnt, &l);
+ 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 (l)
+ // 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, &j, &l);
+ if (ret < 0) {
+ if (!g_allow_regfunc)
+ ferr(po, "non-__cdecl indirect call unhandled yet\n");
+ pp->is_unresolved = 1;
+ j = 0;
+ }
+ j /= 4;
+ if (j > ARRAY_SIZE(pp->arg))
+ ferr(po, "esp adjust too large: %d\n", j);
+ pp->ret_type.name = strdup("int");
+ pp->argc = pp->argc_stack = j;
+ 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
+ ret = -1;
+ if (!pp->is_stdcall && pp->argc_stack > 0)
+ ret = scan_for_esp_adjust(i + 1, opcnt, &j, &l);
+ if (ret >= 0) {
+ if (pp->is_vararg) {
+ if (j / 4 < pp->argc_stack)
+ ferr(po, "esp adjust is too small: %x < %x\n",
+ j, pp->argc_stack * 4);
+ // modify pp to make it have varargs as normal args
+ arg = pp->argc;
+ pp->argc += j / 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 != j / 4)
+ ferr(po, "stack tracking failed for '%s': %x %x\n",
+ tmpname, pp->argc_stack * 4, j);
+
+ ops[ret].flags |= OPF_RMD;
+ 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 if (!l) {
+ // 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",
+ pp->name);
+
+ 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;
+ }
+ }
+
+ // 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
+ // - declare indirect functions
+ for (i = 0; i < opcnt; i++) {
+ po = &ops[i];
+ if (po->flags & OPF_RMD)
+ 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->argnum == 0
+ && !(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) {
+ 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;
+ }
+ }
+ }
+ }
+
+ 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)) {
+ 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) {
+ 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->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;
+
+ if (po->op == OP_RCL || po->op == OP_RCR || po->op == OP_XCHG) {
+ need_tmp_var = 1;
+ }
+ }
+
+ // pass4:
+ // - 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;
+ }