unsigned char pfo;
unsigned char pfo_inv;
unsigned char operand_cnt;
- unsigned char p_argnum; // arg push: altered before call arg #
+ unsigned char p_argnum; // arg push: call's saved arg #
unsigned char p_arggrp; // arg push: arg group # for above
unsigned char p_argpass;// arg push: arg of host func
- short p_argnext;// arg push: same arg pushed elsewhere or -1
+ short pad;
int regmask_src; // all referensed regs
int regmask_dst;
int pfomask; // flagop: parsed_flag_op that can't be delayed
SCTFA_RM_REGS = (1 << 2), // don't emit regs (mask)
SCTFA_NOWARN = (1 << 3), // don't try to detect problems
SCTFA_ARGFRAME = (1 << 4), // copy all args to a struct, in order
+ SCTFA_UA_FLOAT = (1 << 5), // emit float i/o helpers for alignemnt
};
enum x87_const {
SEG_GS,
};
-// note: limited to 32k due to p_argnext
#define MAX_OPS 4096
#define MAX_ARG_GRP 2
const char *prefix = "";
const char *bp_arg = NULL;
char ofs_reg[16] = { 0, };
- char argname[8];
+ char argname[8], buf2[32];
int i, arg_i, arg_s;
int unaligned = 0;
int stack_ra = 0;
arg_i = (offset - stack_ra - 4) / 4;
if (arg_i < 0 || arg_i >= g_func_pp->argc_stack)
{
- if (g_func_pp->is_vararg && arg_i == g_func_pp->argc_stack) {
- if (is_lea) {
+ if (g_func_pp->is_vararg && arg_i >= g_func_pp->argc_stack) {
+ // vararg access - messy and non-portable,
+ // but works with gcc on both x86 and ARM
+ if (arg_i == g_func_pp->argc_stack)
// should be va_list
- if (cast[0] == 0)
- cast = "(u32)";
- snprintf(buf, buf_size, "%sap", cast);
- }
+ snprintf(buf2, sizeof(buf2), "*(u32 *)&ap");
+ else
+ snprintf(buf2, sizeof(buf2), "(*(u32 *)&ap + %u)",
+ (arg_i - g_func_pp->argc_stack) * 4);
+
+ if (is_lea)
+ snprintf(buf, buf_size, "%s%s", cast, buf2);
else
- snprintf(buf, buf_size, "%sva_arg(ap, u32)", cast);
+ snprintf(buf, buf_size, "%s*(u32 *)%s", cast, buf2);
return -1;
}
ferr(po, "offset 0x%x (%s,%d) doesn't map to any arg\n",
return out_src_opr(buf, buf_size, po, popr, NULL, 0);
}
+// do we need a helper func to perform a float i/o?
+static int float_opr_needs_helper(struct parsed_op *po,
+ struct parsed_opr *popr)
+{
+ if (!(g_sct_func_attr & SCTFA_UA_FLOAT))
+ return 0;
+ if (popr->type != OPT_REGMEM)
+ return 0;
+ if (is_stack_access(po, popr))
+ return 0;
+
+ return 1;
+}
+
static char *out_opr_float(char *buf, size_t buf_size,
struct parsed_op *po, struct parsed_opr *popr, int is_src,
int need_float_stack)
break;
}
out_src_opr(tmp, sizeof(tmp), po, popr, "", 1);
- snprintf(buf, buf_size, "*(%s *)(%s)", cast, tmp);
+ if (is_src && float_opr_needs_helper(po, popr))
+ snprintf(buf, buf_size, "%s_load(%s)", cast, tmp);
+ else
+ snprintf(buf, buf_size, "*(%s *)(%s)", cast, tmp);
break;
case OPT_CONST:
const char *s_reg;
int pp_cmp_i;
int arg, reg;
+ int bad = 0;
+ int j;
for (arg = 0; arg < pp->argc; arg++) {
pp_cmp = NULL;
s_opr.reg = reg;
scan_for_call_type(i, &s_opr, i + arg + opcnt * 28, 0,
&pp_cmp, &pp_cmp_i, NULL);
+ if (pp_cmp != NULL && !pp_compatible_func(pp_arg, pp_cmp)) {
+ bad = 1;
+ if (pp_cmp_i >= 0)
+ fnote(&ops[pp_cmp_i], "(referenced here)\n");
+ }
}
else {
- po_a = pp->arg[arg].datap;
- if (po_a != NULL && po_a->op == OP_PUSH)
+ for (j = 0; j < pp->arg[arg].push_ref_cnt; j++) {
+ po_a = pp->arg[arg].push_refs[j];
+ if (po_a == NULL || po_a->op != OP_PUSH)
+ continue;
pp_cmp = resolve_func_ptr(po_a - ops, opcnt, 0,
&po_a->operand[0], &pp_cmp_i, NULL);
- if (pp_cmp_i < 0)
- pp_cmp_i = po_a - ops;
+ if (pp_cmp != NULL && !pp_compatible_func(pp_arg, pp_cmp)) {
+ bad = 1;
+ if (pp_cmp_i < 0)
+ pp_cmp_i = po_a - ops;
+ if (pp_cmp_i >= 0)
+ fnote(&ops[pp_cmp_i], "(referenced here)\n");
+ }
+ }
}
- if (pp_cmp != NULL && !pp_compatible_func(pp_arg, pp_cmp)) {
- if (pp_cmp_i >= 0)
- fnote(&ops[pp_cmp_i], "(referenced here)\n");
+ if (bad)
ferr(&ops[i], "incompatible fptr arg %d\n", arg + 1);
- }
}
}
+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 pp_insert_stack_args(struct parsed_proto *pp, int count)
+{
+ int a;
+
+ pp->argc += count;
+ pp->argc_stack += count;
+
+ for (a = 0; a < pp->argc; a++)
+ if (pp->arg[a].type.name == NULL)
+ pp->arg[a].type.name = strdup("int");
+}
+
+static void pp_add_push_ref(struct parsed_proto *pp,
+ int arg, struct parsed_op *po)
+{
+ pp->arg[arg].push_refs = realloc(pp->arg[arg].push_refs,
+ (pp->arg[arg].push_ref_cnt + 1)
+ * sizeof(pp->arg[arg].push_refs[0]));
+ ferr_assert(po, pp->arg[arg].push_refs != NULL);
+ pp->arg[arg].push_refs[pp->arg[arg].push_ref_cnt++] = po;
+}
+
static void mark_float_arg(struct parsed_op *po,
struct parsed_proto *pp, int arg, int *regmask_ffca)
{
- po->p_argnext = -1;
+ ferr_assert(po, pp->arg[arg].push_ref_cnt == 0);
+ pp_add_push_ref(pp, arg, po);
+
po->p_argnum = arg + 1;
- ferr_assert(po, pp->arg[arg].datap == NULL);
- pp->arg[arg].datap = po;
po->flags |= OPF_DONE | OPF_FARGNR | OPF_FARG;
if (regmask_ffca != NULL)
*regmask_ffca |= 1 << arg;
for (arg = base_arg; arg < pp->argc; arg++) {
ferr_assert(&ops[i], pp->arg[arg].reg == NULL);
- po = pp->arg[arg].datap;
- if (po == NULL)
- ferr(&ops[i], "arg %d/%d not found\n", arg, pp->argc);
+ if (pp->arg[arg].push_ref_cnt != 1)
+ ferr(&ops[i], "arg %d/%d not found or bad\n", arg, pp->argc);
+ po = pp->arg[arg].push_refs[0];
if (po->operand[0].lmod == OPLM_QWORD)
arg++;
}
if (ops[j].op == OP_PUSH)
{
- ops[j].p_argnext = -1;
- ferr_assert(&ops[j], pp->arg[arg].datap == NULL);
+ int ref_handled = 0;
k = check_for_stp(j + 1, i);
if (k != -1) {
if (!pp->arg[arg].type.is_float)
ferr(&ops[i], "arg %d should be float\n", arg + 1);
mark_float_arg(&ops[k], pp, arg, regmask_ffca);
+ ref_handled = 1;
}
}
- if (pp->arg[arg].datap == NULL) {
- pp->arg[arg].datap = &ops[j];
- if (regmask != NULL && ops[j].operand[0].type == OPT_REG)
- *regmask |= 1 << ops[j].operand[0].reg;
+ if (!ref_handled) {
+ ferr_assert(&ops[j], pp->arg[arg].push_ref_cnt == 0);
+ pp_add_push_ref(pp, arg, &ops[j]);
}
+ if (regmask != NULL && 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;
return 0;
}
-static int sync_argnum(struct parsed_op *po, int argnum)
+// ensure all s_a* numbers match for a given func arg in all branches
+// returns 1 if any changes were made, 0 if not
+static int sync_argnum(struct parsed_proto *pp, int arg,
+ int *argnum, int *arggrp)
{
struct parsed_op *po_tmp;
+ int changed = 0;
+ int i;
// see if other branches don't have higher argnum
- for (po_tmp = po; po_tmp != NULL; ) {
- if (argnum < po_tmp->p_argnum)
- argnum = po_tmp->p_argnum;
- // note: p_argnext is active on current collect_call_args only
- po_tmp = po_tmp->p_argnext >= 0 ? &ops[po_tmp->p_argnext] : NULL;
+ for (i = 0; i < pp->arg[arg].push_ref_cnt; i++) {
+ po_tmp = pp->arg[arg].push_refs[i];
+ if (*argnum < po_tmp->p_argnum)
+ *argnum = po_tmp->p_argnum;
+ if (*arggrp < po_tmp->p_arggrp)
+ *arggrp = po_tmp->p_arggrp;
}
// make all argnums consistent
- for (po_tmp = po; po_tmp != NULL; ) {
- if (po_tmp->p_argnum != 0)
- po_tmp->p_argnum = argnum;
- po_tmp = po_tmp->p_argnext >= 0 ? &ops[po_tmp->p_argnext] : NULL;
+ for (i = 0; i < pp->arg[arg].push_ref_cnt; i++) {
+ po_tmp = pp->arg[arg].push_refs[i];
+ if (po_tmp->p_argnum == 0)
+ continue;
+ if (po_tmp->p_argnum != *argnum || po_tmp->p_arggrp != *arggrp) {
+ po_tmp->p_argnum = *argnum;
+ po_tmp->p_arggrp = *arggrp;
+ changed = 1;
+ }
}
- return argnum;
+ return changed;
}
static int collect_call_args_r(struct parsed_op *po, int i,
- struct parsed_proto *pp, int *regmask, int *arg_grp,
- int arg, int argnum, int magic, int need_op_saving, int may_reuse)
+ struct parsed_proto *pp, int *regmask,
+ int arg, int argnum, int magic,
+ int skip, 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 dummy = 0;
int ret = 0;
int reg;
char buf[32];
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, arg_grp,
- arg, argnum, magic, need_op_saving, may_reuse);
+ ret = collect_call_args_r(po, lr->i, pp, regmask,
+ arg, argnum, magic, skip, need_op_saving, may_reuse);
if (ret < 0)
return ret;
}
continue;
}
need_op_saving = 1;
- ret = collect_call_args_r(po, lr->i, pp, regmask, arg_grp,
- arg, argnum, magic, need_op_saving, may_reuse);
+ ret = collect_call_args_r(po, lr->i, pp, regmask,
+ arg, argnum, magic, skip, need_op_saving, may_reuse);
if (ret < 0)
return ret;
}
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);
+ if (!pp_tmp->is_unresolved)
+ skip = 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
may_reuse = 1;
}
+ else if (ops[j].op == OP_PUSH && skip > 0) {
+ // XXX: might want to rm OPF_FARGNR and only use this
+ skip--;
+ }
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];
+ pp_add_push_ref(pp, arg, &ops[j]);
- argnum = sync_argnum(&ops[j], argnum);
+ sync_argnum(pp, arg, &argnum, &dummy);
need_to_save_current = 0;
reg = -1;
if (pp->arg[arg].is_saved) {
ops[j].flags &= ~OPF_RMD;
ops[j].p_argnum = argnum;
+ ops[j].p_arggrp = arg_grp_current;
}
// tracking reg usage
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, int opcnt,
struct parsed_proto *pp, int *regmask, int magic)
{
- // arg group is for cases when pushes for
- // multiple funcs are going on
- struct parsed_op *po_tmp;
- int arg_grp = 0;
int ret;
- int a;
- ret = collect_call_args_r(po, i, pp, regmask, &arg_grp,
- 0, 1, magic, 0, 0);
+ ret = collect_call_args_r(po, i, pp, regmask, 0, 1, magic,
+ 0, 0, 0);
if (ret < 0)
return ret;
- 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");
- }
+ if (pp->is_unresolved)
+ pp_insert_stack_args(pp, 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;
- po_tmp = po_tmp->p_argnext >= 0 ? &ops[po_tmp->p_argnext] : NULL;
- }
- }
- }
+ // note: p_argnum, p_arggrp will be propagated in a later pass,
+ // look for sync_argnum() (p_arggrp is for cases when mixed pushes
+ // for multiple funcs are going on)
if (!g_header_mode)
check_fptr_args(i, opcnt, pp);
}
}
-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 output_std_flag_z(FILE *fout, struct parsed_op *po,
int *pfomask, const char *dst_opr_text)
{
if (pp->is_unresolved) {
int regmask_stack = 0;
- collect_call_args(po, i, opcnt, pp, ®mask, 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 && !pp->arg[arg].is_saved)
- continue;
+ if ((po->flags & OPF_TAIL) && g_func_pp->is_stdcall)
+ pp_insert_stack_args(pp, g_func_pp->argc_stack);
+ else {
+ collect_call_args(po, i, opcnt, pp, ®mask, i + opcnt * 2);
- tmp_op = pp->arg[arg].datap;
- if (tmp_op == NULL)
- ferr(po, "parsed_op missing for arg%d\n", arg);
- if (tmp_op->operand[0].type == OPT_REG)
- regmask_stack |= 1 << tmp_op->operand[0].reg;
+ // 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 && !pp->arg[arg].is_saved)
+ continue;
+
+ if (pp->arg[arg].push_ref_cnt == 0)
+ ferr(po, "parsed_op missing for arg%d\n", arg);
+ tmp_op = pp->arg[arg].push_refs[0];
+ if (tmp_op->operand[0].type == OPT_REG)
+ regmask_stack |= 1 << tmp_op->operand[0].reg;
+ }
}
// quick dumb check for potential reg-args
}
}
- // pass8: final adjustments
+ // pass8: sync all push arg numbers
+ // some calls share args and not all of them
+ // (there's only partial intersection)
+ do {
+ int changed, argnum, arggrp;
+
+ found = 0;
+ for (i = 0; i < opcnt; i++)
+ {
+ po = &ops[i];
+ if ((po->flags & (OPF_RMD|OPF_DONE)) || po->op != OP_CALL)
+ continue;
+
+ pp = po->pp;
+ arggrp = 0;
+ do {
+ changed = 0;
+ for (arg = argnum = 0; arg < pp->argc; arg++) {
+ if (pp->arg[arg].reg != NULL)
+ continue;
+ if (pp->arg[arg].is_saved)
+ changed |= sync_argnum(pp, arg, &argnum, &arggrp);
+ argnum++;
+ }
+ found |= changed;
+ }
+ while (changed);
+
+ if (argnum > 32)
+ ferr(po, "too many args or looping in graph\n");
+ }
+ }
+ while (found);
+
+ // pass9: final adjustments
for (i = 0; i < opcnt; i++)
{
po = &ops[i];
}
// stack arg
- tmp_op = pp->arg[arg].datap;
- if (tmp_op == NULL)
+ if (pp->arg[arg].push_ref_cnt == 0)
ferr(po, "parsed_op missing for arg%d\n", arg);
+ if (pp->arg[arg].push_ref_cnt > 1)
+ ferr_assert(po, pp->arg[arg].is_saved);
+ tmp_op = pp->arg[arg].push_refs[0];
+ ferr_assert(po, tmp_op != NULL);
if (tmp_op->flags & OPF_VAPUSH) {
fprintf(fout, "ap");
ferr_assert(po, cast[0] == 0);
out_src_opr(buf1, sizeof(buf1),
tmp_op, &tmp_op->operand[0], cast, 0);
- tmp_op = pp->arg[++arg].datap;
+ arg++;
+ ferr_assert(po, pp->arg[arg].push_ref_cnt == 1);
+ tmp_op = pp->arg[arg].push_refs[0];
ferr_assert(po, tmp_op != NULL);
out_src_opr(buf2, sizeof(buf2),
tmp_op, &tmp_op->operand[0], cast, 0);
break;
case OP_FST:
+ dead_dst = 0;
if (po->flags & OPF_FARG) {
// store to stack as func arg
- snprintf(buf1, sizeof(buf1), "fs_%d", po->p_argnum);
- dead_dst = 0;
+ fprintf(fout, " fs_%d = %s;", po->p_argnum, float_st0);
+ }
+ else if (po->operand[0].type == OPT_REG
+ && po->operand[0].reg == xST0)
+ {
+ dead_dst = 1;
+ }
+ else if (float_opr_needs_helper(po, &po->operand[0])) {
+ out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], "", 1);
+ fprintf(fout, " %s_store(%s, %s);",
+ po->operand[0].lmod == OPLM_QWORD ? "double" : "float",
+ float_st0, buf1);
}
else {
out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
need_float_stack);
- dead_dst = po->operand[0].type == OPT_REG
- && po->operand[0].reg == xST0;
- }
- if (!dead_dst)
fprintf(fout, " %s = %s;", buf1, float_st0);
+ }
if (po->flags & OPF_FSHIFT) {
if (need_float_stack)
fprintf(fout, " f_stp++;");
"rm_regmask",
"nowarn",
"argframe",
+ "align_float",
};
// parse manual attribute-list comment
notaz.gp2x.de / ia32rtools.git / blobdiff