OP_NOP,
OP_PUSH,
OP_POP,
+ OP_PUSHA,
+ OP_POPA,
OP_LEAVE,
OP_MOV,
OP_LEA,
OP_NOT,
OP_XLAT,
OP_CDQ,
+ OP_BSWAP,
OP_LODS,
OP_STOS,
OP_MOVS,
OP_FILD,
OP_FLDc,
OP_FST,
+ OP_FIST,
OP_FADD,
OP_FDIV,
OP_FMUL,
OP_FISUB,
OP_FIDIVR,
OP_FISUBR,
+ OP_FCOM,
+ OP_FNSTSW,
OP_FCHS,
OP_FCOS,
OP_FPATAN,
// mmx
OP_EMMS,
// pseudo-ops for lib calls
+ OPP_ALLSHL,
+ OPP_ALLSHR,
OPP_FTOL,
// undefined
OP_UD2,
// (OPF_CC) - points to one of (OPF_FLAGS) that affects cc op
// OP_PUSH - points to OP_POP in complex push/pop graph
// OP_POP - points to OP_PUSH in simple push/pop pair
+// OP_FCOM - needed_status_word_bits | (is_z_check << 16)
struct parsed_equ {
char name[64];
}
static const char *parse_stack_el(const char *name, char *extra_reg,
- int early_try)
+ int *base_val, int early_try)
{
const char *p, *p2, *s;
char *endp = NULL;
char buf[32];
- long val;
+ long val = -1;
int len;
if (g_bp_frame || early_try)
if ('0' <= *p && *p <= '9')
return NULL;
+ if (base_val != NULL)
+ *base_val = val;
return p;
}
aerr("[] parse failure\n");
parse_indmode(opr->name, regmask_indirect, 1);
- if (opr->lmod == OPLM_UNSPEC && parse_stack_el(opr->name, NULL, 1))
+ if (opr->lmod == OPLM_UNSPEC
+ && parse_stack_el(opr->name, NULL, NULL, 1))
{
// might be an equ
struct parsed_equ *eq =
- equ_find(NULL, parse_stack_el(opr->name, NULL, 1), &i);
+ equ_find(NULL, parse_stack_el(opr->name, NULL, NULL, 1), &i);
if (eq)
opr->lmod = eq->lmod;
{ "nop", OP_NOP, 0, 0, 0 },
{ "push", OP_PUSH, 1, 1, 0 },
{ "pop", OP_POP, 1, 1, OPF_DATA },
+ { "pusha",OP_PUSHA, 0, 0, 0 },
+ { "popa", OP_POPA, 0, 0, OPF_DATA },
{ "leave",OP_LEAVE, 0, 0, OPF_DATA },
{ "mov" , OP_MOV, 2, 2, OPF_DATA },
{ "lea", OP_LEA, 2, 2, OPF_DATA },
{ "not", OP_NOT, 1, 1, OPF_DATA },
{ "xlat", OP_XLAT, 0, 0, OPF_DATA },
{ "cdq", OP_CDQ, 0, 0, OPF_DATA },
+ { "bswap",OP_BSWAP, 1, 1, OPF_DATA },
{ "lodsb",OP_LODS, 0, 0, OPF_DATA },
{ "lodsw",OP_LODS, 0, 0, OPF_DATA },
{ "lodsd",OP_LODS, 0, 0, OPF_DATA },
{ "fld1", OP_FLDc, 0, 0, OPF_FPUSH },
{ "fldln2", OP_FLDc, 0, 0, OPF_FPUSH },
{ "fldz", OP_FLDc, 0, 0, OPF_FPUSH },
- { "fstp", OP_FST, 1, 1, OPF_FPOP },
{ "fst", OP_FST, 1, 1, 0 },
+ { "fstp", OP_FST, 1, 1, OPF_FPOP },
+ { "fist", OP_FIST, 1, 1, 0 },
+ { "fistp", OP_FIST, 1, 1, OPF_FPOP },
{ "fadd", OP_FADD, 0, 2, 0 },
{ "faddp", OP_FADD, 0, 2, OPF_FPOP },
{ "fdiv", OP_FDIV, 0, 2, 0 },
{ "fisub", OP_FISUB, 1, 1, 0 },
{ "fidivr", OP_FIDIVR, 1, 1, 0 },
{ "fisubr", OP_FISUBR, 1, 1, 0 },
+ { "fcom", OP_FCOM, 0, 1, 0 },
+ { "fcomp", OP_FCOM, 0, 1, OPF_FPOP },
+ { "fnstsw", OP_FNSTSW, 1, 1, OPF_DATA },
{ "fchs", OP_FCHS, 0, 0, 0 },
{ "fcos", OP_FCOS, 0, 0, 0 },
{ "fpatan", OP_FPATAN, 0, 0, OPF_FPOP },
{ "emms", OP_EMMS, 0, 0, OPF_DATA },
{ "movq", OP_MOV, 2, 2, OPF_DATA },
// pseudo-ops for lib calls
+ { "_allshl",OPP_ALLSHL },
+ { "_allshr",OPP_ALLSHR },
{ "_ftol", OPP_FTOL },
// must be last
{ "ud2", OP_UD2 },
case OP_INC:
case OP_DEC:
case OP_NEG:
+ case OP_BSWAP:
// more below..
op->regmask_src |= op->regmask_dst;
break;
break;
case OP_FST:
+ case OP_FIST:
op->regmask_src |= mxST0;
break;
aerr("TODO\n");
break;
+ case OP_FCOM:
+ op->regmask_src |= mxST0;
+ break;
+
default:
break;
}
static int is_stack_access(struct parsed_op *po,
const struct parsed_opr *popr)
{
- return (parse_stack_el(popr->name, NULL, 0)
+ return (parse_stack_el(popr->name, NULL, NULL, 0)
|| (g_bp_frame && !(po->flags & OPF_EBP_S)
&& IS_START(popr->name, "ebp")));
}
ferr(po, "ebp- parse of '%s' failed\n", name);
}
else {
- bp_arg = parse_stack_el(name, ofs_reg, 0);
+ bp_arg = parse_stack_el(name, ofs_reg, NULL, 0);
snprintf(g_comment, sizeof(g_comment), "%s", bp_arg);
eq = equ_find(po, bp_arg, &offset);
if (eq == NULL)
}
*offset_out = offset;
- *stack_ra_out = stack_ra;
+ if (stack_ra_out)
+ *stack_ra_out = stack_ra;
if (bp_arg_out)
*bp_arg_out = bp_arg;
}
+static int parse_stack_esp_offset(struct parsed_op *po,
+ const char *name, int *offset_out)
+{
+ char ofs_reg[16] = { 0, };
+ struct parsed_equ *eq;
+ const char *bp_arg;
+ char *endp = NULL;
+ int base_val = 0;
+ int offset = 0;
+
+ if (strstr(name, "esp") == NULL)
+ return -1;
+ bp_arg = parse_stack_el(name, ofs_reg, &base_val, 0);
+ if (bp_arg == NULL) {
+ // just plain offset?
+ if (!IS_START(name, "esp+"))
+ return -1;
+ offset = strtol(name + 4, &endp, 0);
+ if (endp == NULL || *endp != 0)
+ return -1;
+ *offset_out = offset;
+ return 0;
+ }
+
+ if (ofs_reg[0] != 0)
+ return -1;
+ eq = equ_find(po, bp_arg, &offset);
+ if (eq == NULL)
+ ferr(po, "detected but missing eq\n");
+ offset += eq->offset;
+ *offset_out = base_val + offset;
+ return 0;
+}
+
static int stack_frame_access(struct parsed_op *po,
struct parsed_opr *popr, char *buf, size_t buf_size,
const char *name, const char *cast, int is_src, int is_lea)
}
static void out_cmp_for_cc(char *buf, size_t buf_size,
- struct parsed_op *po, enum parsed_flag_op pfo, int is_inv)
+ struct parsed_op *po, enum parsed_flag_op pfo, int is_inv,
+ int is_neg)
{
const char *cast, *scast, *cast_use;
char buf1[256], buf2[256];
out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], cast_use, 0);
if (po->op == OP_DEC)
snprintf(buf2, sizeof(buf2), "1");
- else
- out_src_opr(buf2, sizeof(buf2), po, &po->operand[1], cast_use, 0);
+ else {
+ char cast_op2[64];
+ snprintf(cast_op2, sizeof(cast_op2) - 1, "%s", cast_use);
+ if (is_neg)
+ strcat(cast_op2, "-");
+ out_src_opr(buf2, sizeof(buf2), po, &po->operand[1], cast_op2, 0);
+ }
switch (pfo) {
case PFO_C:
po->operand[0].lmod, buf3);
}
else if (po->op == OP_CMP) {
- out_cmp_for_cc(buf, buf_size, po, pfo, is_inv);
+ out_cmp_for_cc(buf, buf_size, po, pfo, is_inv, 0);
}
else
ferr(po, "%s: unhandled op: %d\n", __func__, po->op);
unsigned int regmask_src;
unsigned int regmask_dst;
} pseudo_ops[] = {
- { "__ftol", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
+ { "__allshl", OPP_ALLSHL, OPF_DATA, mxAX|mxDX|mxCX, mxAX|mxDX },
+ { "__allshr", OPP_ALLSHR, OPF_DATA, mxAX|mxDX|mxCX, mxAX|mxDX },
+ { "__ftol", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
};
const struct parsed_proto *pp_c;
struct parsed_proto *pp;
tmpname = opr_name(po, 0);
if (IS_START(tmpname, "loc_"))
ferr(po, "call to loc_*\n");
+ if (IS(tmpname, "__alloca_probe"))
+ continue;
// convert some calls to pseudo-ops
for (l = 0; l < ARRAY_SIZE(pseudo_ops); l++) {
static void scan_prologue_epilogue(int opcnt)
{
- int ecx_push = 0, esp_sub = 0;
+ int ecx_push = 0, esp_sub = 0, pusha = 0;
+ int sandard_epilogue;
int found;
int i, j, l;
ops[1].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
i = 2;
- if (ops[2].op == OP_SUB && IS(opr_name(&ops[2], 0), "esp")) {
- g_stack_fsz = opr_const(&ops[2], 1);
- ops[2].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ if (ops[i].op == OP_PUSHA) {
+ ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ pusha = 1;
+ i++;
+ }
+
+ if (ops[i].op == OP_SUB && IS(opr_name(&ops[i], 0), "esp")) {
+ g_stack_fsz = opr_const(&ops[i], 1);
+ ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
i++;
}
else {
// another way msvc builds stack frame..
- i = 2;
while (ops[i].op == OP_PUSH && IS(opr_name(&ops[i], 0), "ecx")) {
g_stack_fsz += 4;
ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
j--;
}
- if ((ops[j].op == OP_POP && IS(opr_name(&ops[j], 0), "ebp"))
- || ops[j].op == OP_LEAVE)
+ sandard_epilogue = 0;
+ if (ops[j].op == OP_POP && IS(opr_name(&ops[j], 0), "ebp"))
+ {
+ ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ // the standard epilogue is sometimes even used without a sf
+ if (ops[j - 1].op == OP_MOV
+ && IS(opr_name(&ops[j - 1], 0), "esp")
+ && IS(opr_name(&ops[j - 1], 1), "ebp"))
+ sandard_epilogue = 1;
+ }
+ else if (ops[j].op == OP_LEAVE)
{
ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ sandard_epilogue = 1;
}
else if (ops[i].op == OP_CALL && ops[i].pp != NULL
&& ops[i].pp->is_noreturn)
else if (!(g_ida_func_attr & IDAFA_NORETURN))
ferr(&ops[j], "'pop ebp' expected\n");
- if (g_stack_fsz != 0) {
+ if (g_stack_fsz != 0 || sandard_epilogue) {
if (ops[j].op == OP_LEAVE)
j--;
- else if (ops[j].op == OP_POP
- && ops[j - 1].op == OP_MOV
- && IS(opr_name(&ops[j - 1], 0), "esp")
- && IS(opr_name(&ops[j - 1], 1), "ebp"))
+ else if (sandard_epilogue) // mov esp, ebp
{
ops[j - 1].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
j -= 2;
}
}
+ if (pusha) {
+ if (ops[j].op == OP_POPA)
+ ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ else
+ ferr(&ops[j], "popa expected\n");
+ }
+
found = 1;
i++;
} while (i < opcnt);
{
g_stack_fsz = ops[i].operand[1].val;
ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ i++;
+ esp_sub = 1;
+ break;
+ }
+ else if (ops[i].op == OP_MOV && ops[i].operand[0].reg == xAX
+ && ops[i].operand[1].type == OPT_CONST
+ && ops[i + 1].op == OP_CALL
+ && IS(opr_name(&ops[i + 1], 0), "__alloca_probe"))
+ {
+ g_stack_fsz += ops[i].operand[1].val;
+ ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ i++;
+ ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ i++;
esp_sub = 1;
break;
}
|| !IS(opr_name(&ops[j], 0), "esp")
|| ops[j].operand[1].type != OPT_CONST
|| ops[j].operand[1].val != g_stack_fsz)
+ {
+ if (ops[i].op == OP_CALL && ops[i].pp != NULL
+ && ops[i].pp->is_noreturn)
+ {
+ // noreturn tailcall with no epilogue
+ i++;
+ continue;
+ }
ferr(&ops[j], "'add esp' expected\n");
+ }
ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
ops[j].operand[1].val = 0; // hack for stack arg scanner
}
}
+// adjust datap of all reachable 'op' insns when moving back
+// returns 1 if at least 1 op was found
+// returns -1 if path without an op was found
+static int adjust_prev_op(int i, enum op_op op, int magic, void *datap)
+{
+ 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 |= adjust_prev_op(lr->i, op, magic, datap);
+ }
+ if (i > 0 && LAST_OP(i - 1))
+ return ret;
+ }
+
+ i--;
+ if (i < 0)
+ return -1;
+
+ if (ops[i].cc_scratch == magic)
+ return 0;
+ ops[i].cc_scratch = magic;
+
+ if (ops[i].op != op)
+ continue;
+
+ ops[i].datap = datap;
+ return 1;
+ }
+}
+
// find next instruction that reads opr
// *op_i must be set to -1 by the caller
// on return, *op_i is set to first referencer insn
}
if (!is_opr_read(opr, po)) {
- if (is_opr_modified(opr, po)
- && (po->op == OP_CALL
- || ((po->flags & OPF_DATA)
- && po->operand[0].lmod == OPLM_DWORD)))
+ int full_opr = 1;
+ if (opr->type == OPT_REG && po->operand[0].type == OPT_REG
+ && opr->reg == po->operand[0].reg && (po->flags & OPF_DATA))
{
+ full_opr = po->operand[0].lmod >= opr->lmod;
+ }
+ if (is_opr_modified(opr, po) && full_opr) {
// it's overwritten
return ret;
}
return 0;
}
+// find next instruction that reads opr
+// *op_i must be set to -1 by the caller
+// on return, *op_i is set to first flag user insn
+// returns 1 if exactly 1 flag user is found
+static int find_next_flag_use(int i, int opcnt, int magic, int *op_i)
+{
+ struct parsed_op *po;
+ int j, ret = 0;
+
+ for (; i < opcnt; i++)
+ {
+ if (ops[i].cc_scratch == magic)
+ return ret;
+ ops[i].cc_scratch = magic;
+
+ po = &ops[i];
+ if (po->op == OP_CALL)
+ return -1;
+ if (po->flags & OPF_JMP) {
+ if (po->btj != NULL) {
+ // jumptable
+ for (j = 0; j < po->btj->count; j++) {
+ check_i(po, po->btj->d[j].bt_i);
+ ret |= find_next_flag_use(po->btj->d[j].bt_i, opcnt,
+ magic, op_i);
+ }
+ return ret;
+ }
+
+ if (po->flags & OPF_RMD)
+ continue;
+ check_i(po, po->bt_i);
+ if (po->flags & OPF_CJMP)
+ goto found;
+ else
+ i = po->bt_i - 1;
+ continue;
+ }
+
+ if (!(po->flags & OPF_CC)) {
+ if (po->flags & OPF_FLAGS)
+ // flags changed
+ return ret;
+ if (po->flags & OPF_TAIL)
+ return ret;
+ continue;
+ }
+
+found:
+ if (*op_i >= 0)
+ return -1;
+
+ *op_i = i;
+ return 1;
+ }
+
+ return 0;
+}
+
static int try_resolve_const(int i, const struct parsed_opr *opr,
int magic, unsigned int *val)
{
return -1;
}
+static int resolve_used_bits(int i, int opcnt, int reg,
+ int *mask, int *is_z_check)
+{
+ struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_WORD, reg);
+ int j = -1, k = -1;
+ int ret;
+
+ ret = find_next_read(i, opcnt, &opr, i + opcnt * 20, &j);
+ if (ret != 1)
+ return -1;
+
+ find_next_read(j + 1, opcnt, &opr, i + opcnt * 20 + 1, &k);
+ if (k != -1) {
+ fnote(&ops[j], "(first read)\n");
+ ferr(&ops[k], "TODO: bit resolve: multiple readers\n");
+ }
+
+ if (ops[j].op != OP_TEST || ops[j].operand[1].type != OPT_CONST)
+ ferr(&ops[j], "TODO: bit resolve: not a const test\n");
+
+ ferr_assert(&ops[j], ops[j].operand[0].type == OPT_REG);
+ ferr_assert(&ops[j], ops[j].operand[0].reg == reg);
+
+ *mask = ops[j].operand[1].val;
+ if (ops[j].operand[0].lmod == OPLM_BYTE
+ && ops[j].operand[0].name[1] == 'h')
+ {
+ *mask <<= 8;
+ }
+ ferr_assert(&ops[j], (*mask & ~0xffff) == 0);
+
+ *is_z_check = 0;
+ ret = find_next_flag_use(j + 1, opcnt, i + opcnt * 20 + 2, &k);
+ if (ret == 1)
+ *is_z_check = ops[k].pfo == PFO_Z;
+
+ return 0;
+}
+
static const struct parsed_proto *resolve_icall(int i, int opcnt,
int *pp_i, int *multi_src)
{
return pp;
}
-static int collect_call_args_early(struct parsed_op *po, int i,
- struct parsed_proto *pp, int *regmask)
+static int collect_call_args_no_push(int i, struct parsed_proto *pp,
+ int *regmask_ffca)
{
+ struct parsed_op *po;
+ int offset = 0;
+ int base_arg;
+ int j, arg;
+ int ret;
+
+ for (base_arg = 0; base_arg < pp->argc; base_arg++)
+ if (pp->arg[base_arg].reg == NULL)
+ break;
+
+ for (j = i; j > 0; )
+ {
+ ferr_assert(&ops[j], g_labels[j] == NULL);
+ j--;
+
+ po = &ops[j];
+ ferr_assert(po, po->op != OP_PUSH);
+ if (po->op == OP_FST)
+ {
+ if (po->operand[0].type != OPT_REGMEM)
+ continue;
+ ret = parse_stack_esp_offset(po, po->operand[0].name, &offset);
+ if (ret != 0)
+ continue;
+ if (offset < 0 || offset >= pp->argc_stack * 4 || (offset & 3))
+ ferr(po, "bad offset %d (%d args)\n", offset, pp->argc_stack);
+
+ arg = base_arg + offset / 4;
+ po->p_argnext = -1;
+ 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;
+ }
+ else if (po->op == OP_SUB && po->operand[0].reg == xSP
+ && po->operand[1].type == OPT_CONST)
+ {
+ po->flags |= OPF_RMD | OPF_DONE | OPF_FARGNR | OPF_FARG;
+ break;
+ }
+ }
+
+ 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 (po->operand[0].lmod == OPLM_QWORD)
+ arg++;
+ }
+
+ return 0;
+}
+
+static int collect_call_args_early(int i, struct parsed_proto *pp,
+ int *regmask, int *regmask_ffca)
+{
+ struct parsed_op *po;
int arg, ret;
int j;
return -1;
j--;
- if (ops[j].op == OP_CALL)
+ po = &ops[j];
+ if (po->op == OP_CALL)
return -1;
- else if (ops[j].op == OP_ADD && ops[j].operand[0].reg == xSP)
+ else if (po->op == OP_ADD && po->operand[0].reg == xSP)
return -1;
- else if (ops[j].op == OP_POP)
+ else if (po->op == OP_POP)
return -1;
- else if (ops[j].flags & OPF_CJMP)
+ else if (po->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)
+ else if (po->op == OP_PUSH) {
+ if (po->flags & (OPF_FARG|OPF_FARGNR))
return -1;
+ if (!g_header_mode) {
+ ret = scan_for_mod(po, j + 1, i, 1);
+ if (ret >= 0)
+ return -1;
+ }
if (pp->arg[arg].type.is_va_list)
return -1;
if (pp->arg[arg].reg == NULL)
break;
}
+ else if (po->op == OP_SUB && po->operand[0].reg == xSP
+ && po->operand[1].type == OPT_CONST)
+ {
+ if (po->flags & (OPF_RMD|OPF_DONE))
+ return -1;
+ if (po->operand[1].val != pp->argc_stack * 4)
+ ferr(po, "unexpected esp adjust: %d\n",
+ po->operand[1].val * 4);
+ ferr_assert(po, pp->argc - arg == pp->argc_stack);
+ return collect_call_args_no_push(i, pp, regmask_ffca);
+ }
}
if (arg < pp->argc)
ferr_assert(&ops[j], pp->arg[arg].datap == NULL);
pp->arg[arg].datap = &ops[j];
- if (ops[j].operand[0].type == OPT_REG)
+ 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;
if (po->flags & OPF_NOREGS)
continue;
- if (po->flags & OPF_FPUSH) {
- if (regmask_now & mxST1)
- regmask_now |= mxSTa; // switch to "full stack" mode
- if (regmask_now & mxSTa)
- po->flags |= OPF_FSHIFT;
- if (!(regmask_now & mxST7_2)) {
- regmask_now =
- (regmask_now & ~mxST1_0) | ((regmask_now & mxST0) << 1);
- }
- }
-
// if incomplete register is used, clear it on init to avoid
// later use of uninitialized upper part in some situations
if ((po->flags & OPF_DATA) && po->operand[0].type == OPT_REG
}
}
+ if (po->flags & OPF_FPUSH) {
+ if (regmask_now & mxST1)
+ regmask_now |= mxSTa; // switch to "full stack" mode
+ if (regmask_now & mxSTa)
+ po->flags |= OPF_FSHIFT;
+ if (!(regmask_now & mxST7_2)) {
+ regmask_now =
+ (regmask_now & ~mxST1_0) | ((regmask_now & mxST0) << 1);
+ }
+ }
+
regmask_now |= regmask_op;
*regmask |= regmask_now;
}
if (po->flags & OPF_TAIL) {
- if (!(regmask_now & mxST7_2) && (regmask_now & mxST1_0))
- ferr(po, "float regs on tail: %x\n", regmask_now);
+ if (!(regmask_now & mxST7_2)) {
+ if (get_pp_arg_regmask_dst(g_func_pp) & mxST0) {
+ if (!(regmask_now & mxST0))
+ ferr(po, "no st0 on float return, mask: %x\n",
+ regmask_now);
+ }
+ else if (regmask_now & mxST1_0)
+ ferr(po, "float regs on tail: %x\n", regmask_now);
+ }
// there is support for "conditional tailcall", sort of
if (!(po->flags & OPF_CC))
const char *float_st0;
const char *float_st1;
int need_float_stack = 0;
+ int need_float_sw = 0; // status word
int need_tmp_var = 0;
int need_tmp64 = 0;
int cond_vars = 0;
int had_decl = 0;
int label_pending = 0;
int need_double = 0;
- int regmask_save = 0; // regs saved/restored in this func
+ int regmask_save = 0; // used regs saved/restored in this func
int regmask_arg; // regs from this function args (fastcall, etc)
int regmask_ret; // regs needed on ret
int regmask_now; // temp
int regmask_init = 0; // regs that need zero initialization
int regmask_pp = 0; // regs used in complex push-pop graph
+ int regmask_ffca = 0; // float function call args
int regmask = 0; // used regs
int pfomask = 0;
int found = 0;
{
pp = process_call_early(i, opcnt, &j);
if (pp != NULL) {
- if (!(po->flags & OPF_ATAIL))
+ if (!(po->flags & OPF_ATAIL)) {
// since we know the args, try to collect them
- if (collect_call_args_early(po, i, pp, ®mask) != 0)
+ ret = collect_call_args_early(i, pp, ®mask, ®mask_ffca);
+ if (ret != 0)
pp = NULL;
+ }
}
if (pp != NULL) {
// - process calls, stage 2
// - handle some push/pop pairs
// - scan for STD/CLD, propagate DF
+ // - try to resolve needed x87 status word bits
for (i = 0; i < opcnt; i++)
{
+ int mask, z_check;
+
po = &ops[i];
if (po->flags & OPF_RMD)
continue;
if (po->flags & OPF_DONE)
continue;
- if (po->op == OP_PUSH && !(po->flags & OPF_FARG)
- && !(po->flags & OPF_RSAVE) && po->operand[0].type == OPT_CONST)
- {
- scan_for_pop_const(i, opcnt, i + opcnt * 12);
- }
- else if (po->op == OP_POP)
+ switch (po->op) {
+ case OP_PUSH:
+ if (!(po->flags & OPF_FARG) && !(po->flags & OPF_RSAVE)
+ && po->operand[0].type == OPT_CONST)
+ {
+ scan_for_pop_const(i, opcnt, i + opcnt * 12);
+ }
+ break;
+
+ case OP_POP:
scan_pushes_for_pop(i, opcnt, ®mask_pp);
- else if (po->op == OP_STD) {
+ break;
+
+ case OP_STD:
po->flags |= OPF_DF | OPF_RMD | OPF_DONE;
scan_propagate_df(i + 1, opcnt);
+ break;
+
+ case OP_FNSTSW:
+ need_float_sw = 1;
+ if (po->operand[0].type != OPT_REG || po->operand[0].reg != xAX)
+ ferr(po, "TODO: fnstsw to mem\n");
+ ret = resolve_used_bits(i + 1, opcnt, xAX, &mask, &z_check);
+ if (ret != 0)
+ ferr(po, "fnstsw resolve failed\n");
+ ret = adjust_prev_op(i, OP_FCOM, i + opcnt * 21,
+ (void *)(long)(mask | (z_check << 16)));
+ if (ret != 1)
+ ferr(po, "failed to find fcom: %d\n", ret);
+ break;
+
+ default:
+ break;
}
}
// - find flag set ops for their users
// - do unresolved calls
// - declare indirect functions
+ // - other op specific processing
for (i = 0; i < opcnt; i++)
{
po = &ops[i];
cond_vars |= 1 << PFO_C;
}
- if (po->op == OP_CMPS || po->op == OP_SCAS) {
+ switch (po->op) {
+ case OP_CMPS:
+ case OP_SCAS:
cond_vars |= 1 << PFO_Z;
- }
- else if (po->op == OP_MUL
- || (po->op == OP_IMUL && po->operand_cnt == 1))
- {
+ break;
+
+ case OP_MUL:
if (po->operand[0].lmod == OPLM_DWORD)
need_tmp64 = 1;
- }
- else if (po->op == OP_CALL) {
+ break;
+
+ case OP_IMUL:
+ if (po->operand_cnt == 1 && po->operand[0].lmod == OPLM_DWORD)
+ need_tmp64 = 1;
+ break;
+
+ case OP_CALL:
// note: resolved non-reg calls are OPF_DONE already
pp = po->pp;
ferr_assert(po, pp != NULL);
if (pp->argc_stack > 0)
pp->is_stdcall = 1;
}
- }
- 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_"))
+ break;
+
+ case OP_MOV:
+ if (po->operand[0].pp != NULL && po->operand[1].pp != NULL)
{
- 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");
+ // <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_DIV || po->op == OP_IDIV) {
+ break;
+
+ case OP_DIV:
+ case OP_IDIV:
if (po->operand[0].lmod == OPLM_DWORD) {
// 32bit division is common, look for it
if (po->op == OP_DIV)
}
else
need_tmp_var = 1;
- }
- else if (po->op == OP_CLD)
+ break;
+
+ case OP_CLD:
po->flags |= OPF_RMD | OPF_DONE;
- else if (po->op == OPP_FTOL) {
+ break;
+
+ case OP_RCL:
+ case OP_RCR:
+ case OP_XCHG:
+ need_tmp_var = 1;
+ break;
+
+ case OP_FLD:
+ if (po->operand[0].lmod == OPLM_QWORD)
+ need_double = 1;
+ break;
+
+ case OPP_ALLSHL:
+ case OPP_ALLSHR:
+ need_tmp64 = 1;
+ break;
+
+ case OPP_FTOL: {
struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, xDX);
j = -1;
find_next_read(i + 1, opcnt, &opr, i + opcnt * 18, &j);
if (j == -1)
po->flags |= OPF_32BIT;
+ break;
}
- else if (po->op == OP_FLD && po->operand[0].lmod == OPLM_QWORD)
- need_double = 1;
- if (po->op == OP_RCL || po->op == OP_RCR || po->op == OP_XCHG)
- need_tmp_var = 1;
+ default:
+ break;
+ }
}
float_type = need_double ? "double" : "float";
}
}
+ if (need_float_sw) {
+ fprintf(fout, " u16 f_sw;\n");
+ had_decl = 1;
+ }
+
if (regmask_save) {
for (reg = 0; reg < 8; reg++) {
if (regmask_save & (1 << reg)) {
}
}
+ if (regmask_ffca) {
+ for (reg = 0; reg < 32; reg++) {
+ if (regmask_ffca & (1 << reg)) {
+ fprintf(fout, " %s fs_%d;\n", float_type, reg + 1);
+ had_decl = 1;
+ }
+ }
+ }
+
// declare push-pop temporaries
if (regmask_pp) {
for (reg = 0; reg < 8; reg++) {
strcpy(g_comment, "cdq");
break;
+ case OP_BSWAP:
+ assert_operand_cnt(1);
+ out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+ fprintf(fout, " %s = __builtin_bswap32(%s);", buf1, buf1);
+ break;
+
case OP_LODS:
if (po->flags & OPF_REP) {
assert_operand_cnt(3);
delayed_flag_op = NULL;
break;
}
+ if (pfomask & (1 << PFO_LE)) {
+ out_cmp_for_cc(buf1, sizeof(buf1), po, PFO_LE, 0, 1);
+ fprintf(fout, " cond_%s = %s;\n",
+ parsed_flag_op_names[PFO_LE], buf1);
+ pfomask &= ~(1 << PFO_LE);
+ }
goto dualop_arith;
case OP_SUB:
if (j == PFO_Z || j == PFO_S)
continue;
- out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0);
+ out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0, 0);
fprintf(fout, " cond_%s = %s;\n",
parsed_flag_op_names[j], buf1);
pfomask &= ~(1 << j);
if (j == PFO_Z || j == PFO_S || j == PFO_C)
continue;
- out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0);
+ out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0, 0);
fprintf(fout, " cond_%s = %s;\n",
parsed_flag_op_names[j], buf1);
pfomask &= ~(1 << j);
if (tmp_op->flags & OPF_VAPUSH) {
fprintf(fout, "ap");
}
+ else if (tmp_op->op == OP_FST) {
+ fprintf(fout, "fs_%d", tmp_op->p_argnum);
+ if (tmp_op->operand[0].lmod == OPLM_QWORD)
+ arg++;
+ }
else if (tmp_op->p_argpass != 0) {
fprintf(fout, "a%d", tmp_op->p_argpass);
}
g_func_pp->arg[arg].reg, g_func_pp->arg[arg].reg);
}
- if (!(regmask_ret & (1 << xAX))) {
+ if (regmask_ret & mxST0) {
+ fprintf(fout, " return %s;", float_st0);
+ }
+ else if (!(regmask_ret & mxAX)) {
if (i != opcnt - 1 || label_pending)
fprintf(fout, " return;");
}
no_output = 1;
break;
+ // pseudo ops
+ case OPP_ALLSHL:
+ case OPP_ALLSHR:
+ fprintf(fout, " tmp64 = ((u64)edx << 32) | eax;\n");
+ fprintf(fout, " tmp64 = (s64)tmp64 %s= LOBYTE(ecx);\n",
+ po->op == OPP_ALLSHL ? "<<" : ">>");
+ fprintf(fout, " edx = tmp64 >> 32; eax = tmp64;");
+ strcat(g_comment, po->op == OPP_ALLSHL
+ ? " allshl" : " allshr");
+ break;
+
// x87
case OP_FLD:
if (need_float_stack) {
break;
case OP_FST:
- 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 (need_float_stack) {
- if (!dead_dst)
- fprintf(fout, " %s = f_st[f_stp & 7];", buf1);
- if (po->flags & OPF_FSHIFT)
- fprintf(fout, " f_stp++;");
+ if (po->flags & OPF_FARG) {
+ // store to stack as func arg
+ snprintf(buf1, sizeof(buf1), "fs_%d", po->p_argnum);
+ dead_dst = 0;
}
else {
- if (!dead_dst)
- fprintf(fout, " %s = f_st0;", buf1);
- if (po->flags & OPF_FSHIFT)
+ 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++;");
+ else
fprintf(fout, " f_st0 = f_st1;");
}
if (dead_dst && !(po->flags & OPF_FSHIFT))
strcat(g_comment, " fst");
break;
+ case OP_FIST:
+ fprintf(fout, " %s = %s%s;",
+ out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+ lmod_cast(po, po->operand[0].lmod, 1), float_st0);
+ if (po->flags & OPF_FSHIFT) {
+ if (need_float_stack)
+ fprintf(fout, " f_stp++;");
+ else
+ fprintf(fout, " f_st0 = f_st1;");
+ }
+ strcat(g_comment, " fist");
+ break;
+
case OP_FADD:
case OP_FDIV:
case OP_FMUL:
po->op == OP_FIDIVR ? '/' : '-', float_st0);
break;
+ case OP_FCOM: {
+ int mask, z_check;
+ ferr_assert(po, po->datap != NULL);
+ mask = (long)po->datap & 0xffff;
+ z_check = ((long)po->datap >> 16) & 1;
+ out_src_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
+ need_float_stack);
+ if (mask == 0x0100) { // C0 -> <
+ fprintf(fout, " f_sw = %s < %s ? 0x0100 : 0;",
+ float_st0, buf1);
+ }
+ else if (mask == 0x4000) { // C3 -> =
+ fprintf(fout, " f_sw = %s == %s ? 0x4000 : 0;",
+ float_st0, buf1);
+ }
+ else if (mask == 0x4100) { // C3, C0
+ if (z_check) {
+ fprintf(fout, " f_sw = %s <= %s ? 0x4100 : 0;",
+ float_st0, buf1);
+ strcat(g_comment, " z_chk_det");
+ }
+ else {
+ fprintf(fout, " f_sw = %s == %s ? 0x4000 : "
+ "(%s < %s ? 0x0100 : 0);",
+ float_st0, buf1, float_st0, buf1);
+ }
+ }
+ else
+ ferr(po, "unhandled sw mask: %x\n", mask);
+ if (po->flags & OPF_FSHIFT) {
+ if (need_float_stack)
+ fprintf(fout, " f_stp++;");
+ else
+ fprintf(fout, " f_st0 = f_st1;");
+ }
+ break;
+ }
+
+ case OP_FNSTSW:
+ fprintf(fout, " %s = f_sw;",
+ out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]));
+ break;
+
case OP_FCHS:
fprintf(fout, " %s = -%s;", float_st0, float_st0);
break;
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_dummy) != 0)
+ if (collect_call_args_early(i, pp, NULL, NULL) != 0)
pp = NULL;
}
fp->argc_stack--;
}
- fp->regmask_dep = regmask_dep & ~(1 << xSP);
+ fp->regmask_dep = regmask_dep & ~((1 << xSP) | mxSTa);
fp->has_ret = has_ret;
#if 0
printf("// has_ret %d, regmask_dep %x\n",
notaz.gp2x.de / ia32rtools.git / blobdiff