*
* This work is licensed under the terms of 3-clause BSD license.
* See COPYING file in the top-level directory.
+ *
+ * recognized asm hint comments:
+ * sctattr - function attributes (see code)
+ * sctend - force end of function/chunk
+ * sctpatch: <p> - replace current asm line with <p>
+ * sctproto: <p> - prototype of ref'd function or struct
+ * sctref - variable is referenced, make global
+ * sctskip_start - start of skipped code chunk (inclusive)
+ * sctskip_end - end of skipped code chunk (inclusive)
*/
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include <errno.h>
#include "my_assert.h"
#include "my_str.h"
#include "common.h"
-#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
-#define IS(w, y) !strcmp(w, y)
-#define IS_START(w, y) !strncmp(w, y, strlen(y))
-
#include "protoparse.h"
static const char *asmfn;
OPF_FPUSH = (1 << 22), /* pushes x87 stack */
OPF_FPOP = (1 << 23), /* pops x87 stack */
OPF_FSHIFT = (1 << 24), /* x87 stack shift is actually needed */
+ OPF_FINT = (1 << 25), /* integer float op arg */
};
enum op_op {
OP_ADC,
OP_SBB,
OP_BSF,
+ OP_BSR,
OP_INC,
OP_DEC,
OP_NEG,
OPP_ALLSHL,
OPP_ALLSHR,
OPP_FTOL,
+ OPP_CIPOW,
+ OPP_ABORT,
// undefined
OP_UD2,
};
static int g_regmask_init;
static int g_skip_func;
static int g_allow_regfunc;
+static int g_allow_user_icall;
static int g_quiet_pp;
static int g_header_mode;
#define mxAX (1 << xAX)
#define mxCX (1 << xCX)
#define mxDX (1 << xDX)
+#define mxSP (1 << xSP)
#define mxST0 (1 << xST0)
#define mxST1 (1 << xST1)
#define mxST1_0 (mxST1 | mxST0)
}
if ('0' <= w[0] && w[0] <= '9') {
- number = parse_number(w);
+ number = parse_number(w, 0);
printf_number(d, sizeof(cvtbuf) - (d - cvtbuf), number);
continue;
}
if (len < sizeof(buf) - 1) {
strncpy(buf, s, len);
buf[len] = 0;
+ errno = 0;
val = strtol(buf, &endp, 16);
- if (val == 0 || *endp != 0) {
+ if (val == 0 || *endp != 0 || errno != 0) {
aerr("%s num parse fail for '%s'\n", __func__, buf);
return NULL;
}
static int guess_lmod_from_c_type(enum opr_lenmod *lmod,
const struct parsed_type *c_type)
{
+ static const char *qword_types[] = {
+ "uint64_t", "int64_t", "__int64",
+ };
static const char *dword_types[] = {
"uint32_t", "int", "_DWORD", "UINT_PTR", "DWORD",
"WPARAM", "LPARAM", "UINT", "__int32",
}
}
+ for (i = 0; i < ARRAY_SIZE(qword_types); i++) {
+ if (IS(n, qword_types[i])) {
+ *lmod = OPLM_QWORD;
+ return 1;
+ }
+ }
+
return 0;
}
else if (('0' <= words[w][0] && words[w][0] <= '9')
|| words[w][0] == '-')
{
- number = parse_number(words[w]);
+ number = parse_number(words[w], 0);
opr->type = OPT_CONST;
opr->val = number;
printf_number(opr->name, sizeof(opr->name), number);
{ "adc", OP_ADC, 2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
{ "sbb", OP_SBB, 2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
{ "bsf", OP_BSF, 2, 2, OPF_DATA|OPF_FLAGS },
+ { "bsr", OP_BSR, 2, 2, OPF_DATA|OPF_FLAGS },
{ "inc", OP_INC, 1, 1, OPF_DATA|OPF_FLAGS },
{ "dec", OP_DEC, 1, 1, OPF_DATA|OPF_FLAGS },
{ "neg", OP_NEG, 1, 1, OPF_DATA|OPF_FLAGS },
{ "setnle", OP_SCC, 1, 1, OPF_DATA|OPF_CC, PFO_LE, 1 },
// x87
{ "fld", OP_FLD, 1, 1, OPF_FPUSH },
- { "fild", OP_FILD, 1, 1, OPF_FPUSH },
+ { "fild", OP_FILD, 1, 1, OPF_FPUSH|OPF_FINT },
{ "fld1", OP_FLDc, 0, 0, OPF_FPUSH },
{ "fldln2", OP_FLDc, 0, 0, OPF_FPUSH },
{ "fldz", OP_FLDc, 0, 0, OPF_FPUSH },
{ "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 },
+ { "fist", OP_FIST, 1, 1, OPF_FINT },
+ { "fistp", OP_FIST, 1, 1, OPF_FPOP|OPF_FINT },
{ "fadd", OP_FADD, 0, 2, 0 },
{ "faddp", OP_FADD, 0, 2, OPF_FPOP },
{ "fdiv", OP_FDIV, 0, 2, 0 },
{ "fdivrp", OP_FDIVR, 0, 2, OPF_FPOP },
{ "fsubr", OP_FSUBR, 0, 2, 0 },
{ "fsubrp", OP_FSUBR, 0, 2, OPF_FPOP },
- { "fiadd", OP_FIADD, 1, 1, 0 },
- { "fidiv", OP_FIDIV, 1, 1, 0 },
- { "fimul", OP_FIMUL, 1, 1, 0 },
- { "fisub", OP_FISUB, 1, 1, 0 },
- { "fidivr", OP_FIDIVR, 1, 1, 0 },
- { "fisubr", OP_FISUBR, 1, 1, 0 },
+ { "fiadd", OP_FIADD, 1, 1, OPF_FINT },
+ { "fidiv", OP_FIDIV, 1, 1, OPF_FINT },
+ { "fimul", OP_FIMUL, 1, 1, OPF_FINT },
+ { "fisub", OP_FISUB, 1, 1, OPF_FINT },
+ { "fidivr", OP_FIDIVR, 1, 1, OPF_FINT },
+ { "fisubr", OP_FISUBR, 1, 1, OPF_FINT },
{ "fcom", OP_FCOM, 0, 1, 0 },
{ "fcomp", OP_FCOM, 0, 1, OPF_FPOP },
{ "fnstsw", OP_FNSTSW, 1, 1, OPF_DATA },
{ "_allshl",OPP_ALLSHL },
{ "_allshr",OPP_ALLSHR },
{ "_ftol", OPP_FTOL },
+ { "_CIpow", OPP_CIPOW },
+ { "abort", OPP_ABORT },
// must be last
{ "ud2", OP_UD2 },
};
*extra_offs = 0;
namelen = strlen(name);
- p = strchr(name, '+');
+ p = strpbrk(name, "+-");
if (p != NULL) {
namelen = p - name;
if (namelen <= 0)
ferr(po, "equ parse failed for '%s'\n", name);
- if (IS_START(p, "0x"))
- p += 2;
+ errno = 0;
*extra_offs = strtol(p, &endp, 16);
- if (*endp != 0)
+ if (*endp != 0 || errno != 0)
ferr(po, "equ parse failed for '%s'\n", name);
}
p = name + 4;
if (IS_START(p, "0x"))
p += 2;
+ errno = 0;
offset = strtoul(p, &endp, 16);
if (name[3] == '-')
offset = -offset;
- if (*endp != 0)
+ if (*endp != 0 || errno != 0)
ferr(po, "ebp- parse of '%s' failed\n", name);
}
else {
// just plain offset?
if (!IS_START(name, "esp+"))
return -1;
+ errno = 0;
offset = strtol(name + 4, &endp, 0);
- if (endp == NULL || *endp != 0)
+ if (endp == NULL || *endp != 0 || errno != 0)
return -1;
*offset_out = offset;
return 0;
int offset = 0;
int retval = -1;
int sf_ofs;
- int lim;
- if (po->flags & OPF_EBP_S)
+ if (g_bp_frame && (po->flags & OPF_EBP_S)
+ && !(po->regmask_src & mxSP))
ferr(po, "stack_frame_access while ebp is scratch\n");
parse_stack_access(po, name, ofs_reg, &offset,
}
break;
+ case OPLM_QWORD:
+ ferr_assert(po, !(offset & 7));
+ if (cast[0])
+ prefix = cast;
+ snprintf(buf, buf_size, "%s%sa%d",
+ prefix, is_lea ? "&" : "", i + 1);
+ break;
+
default:
ferr(po, "bp_arg bad lmod: %d\n", popr->lmod);
}
g_stack_frame_used = 1;
sf_ofs = g_stack_fsz + offset;
- lim = (ofs_reg[0] != 0) ? -4 : 0;
- if (offset > 0 || sf_ofs < lim)
+ if (ofs_reg[0] == 0 && (offset > 0 || sf_ofs < 0))
ferr(po, "bp_stack offset %d/%d\n", offset, g_stack_fsz);
if (is_lea)
case OPLM_QWORD:
ferr_assert(po, !(sf_ofs & 7));
ferr_assert(po, ofs_reg[0] == 0);
- // float callers set is_lea
- ferr_assert(po, is_lea);
+ // only used for x87 int64/float, float sets is_lea
+ if (!is_lea && (po->flags & OPF_FINT))
+ prefix = "*(s64 *)&";
snprintf(buf, buf_size, "%ssf.q[%d]", prefix, sf_ofs / 8);
break;
int ret, i;
if (pp->argc_reg != 0) {
- if (/*!g_allow_regfunc &&*/ !pp->is_fastcall) {
+ if (!g_allow_user_icall && !pp->is_fastcall) {
pp_print(buf, sizeof(buf), pp);
ferr(po, "%s: unexpected reg arg in icall: %s\n", pfx, buf);
}
return out_src_opr(buf, buf_size, po, popr, NULL, 0);
}
-static char *out_src_opr_float(char *buf, size_t buf_size,
- struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
+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)
{
const char *cast = NULL;
char tmp[256];
break;
case OPT_REGMEM:
+ if (popr->lmod == OPLM_QWORD && is_stack_access(po, popr)) {
+ stack_frame_access(po, popr, buf, buf_size,
+ popr->name, "", is_src, 0);
+ break;
+ }
+ // fallthrough
case OPT_LABEL:
case OPT_OFFSET:
switch (popr->lmod) {
break;
}
out_src_opr(tmp, sizeof(tmp), po, popr, "", 1);
- snprintf(buf, buf_size, "*((%s *)%s)", cast, tmp);
+ snprintf(buf, buf_size, "*(%s *)(%s)", cast, tmp);
break;
default:
return buf;
}
+static char *out_src_opr_float(char *buf, size_t buf_size,
+ struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
+{
+ return out_opr_float(buf, buf_size, po, popr, 1, need_float_stack);
+}
+
static char *out_dst_opr_float(char *buf, size_t buf_size,
struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
{
- // same?
- return out_src_opr_float(buf, buf_size, po, popr, need_float_stack);
+ return out_opr_float(buf, buf_size, po, popr, 0, need_float_stack);
}
static void out_test_for_cc(char *buf, size_t buf_size,
switch (pfo) {
case PFO_Z:
- case PFO_BE: // CF=1||ZF=1; CF=0
+ case PFO_BE: // CF==1||ZF==1; CF=0
snprintf(buf, buf_size, "(%s%s %s 0)",
cast, expr, is_inv ? "!=" : "==");
break;
scast, expr, is_inv ? ">=" : "<");
break;
- case PFO_LE: // ZF=1||SF!=OF; OF=0
+ case PFO_LE: // ZF==1||SF!=OF; OF=0
snprintf(buf, buf_size, "(%s%s %s 0)",
scast, expr, is_inv ? ">" : "<=");
break;
+ case PFO_C: // CF=0
+ case PFO_O: // OF=0
+ snprintf(buf, buf_size, "(%d)", !!is_inv);
+ break;
+
+ case PFO_P: // PF==1
+ snprintf(buf, buf_size, "(%sdo_parity(%s))",
+ is_inv ? "!" : "", expr);
+ break;
+
default:
ferr(po, "%s: unhandled parsed_flag_op: %d\n", __func__, pfo);
}
if (po->pp != NULL && po->pp->is_noreturn)
seen_noreturn = 1;
else
- return -1;
+ goto out;
}
else
- return -1; // deadend
+ goto out;
}
if (po->flags & (OPF_RMD|OPF_DONE|OPF_FARG))
}
}
+out:
// for noreturn, assume msvc skipped stack cleanup
return seen_noreturn ? 1 : -1;
}
return -1;
}
- // nothing interesting on this path
- return 0;
+ // nothing interesting on this path,
+ // still return ret for something recursive calls could find
+ return ret;
}
static void find_reachable_exits(int i, int opcnt, int magic,
{
static int exits[MAX_EXITS];
static int exit_count;
- int j, ret;
+ int found = 0;
+ int e, j, ret;
if (!set_flags) {
exit_count = 0;
}
for (j = 0; j < exit_count; j++) {
- ret = scan_for_rsave_pop_reg(exits[j], i + opcnt * 16 + set_flags,
+ e = exits[j];
+ ret = scan_for_rsave_pop_reg(e, i + opcnt * 16 + set_flags,
reg, set_flags);
- if (ret == -1)
- return -1;
+ if (ret != -1) {
+ found |= ret;
+ continue;
+ }
+ if (ops[e].op == OP_CALL && ops[e].pp != NULL
+ && ops[e].pp->is_noreturn)
+ {
+ // assume stack cleanup was skipped
+ continue;
+ }
+ return -1;
}
- return 1;
+ return found;
}
// scan for one or more pop of push <const>
if (*pp_found != NULL && pp != NULL && *pp_found != pp) {
if (!IS((*pp_found)->ret_type.name, pp->ret_type.name)
|| (*pp_found)->is_stdcall != pp->is_stdcall
- || (*pp_found)->is_fptr != pp->is_fptr
+ //|| (*pp_found)->is_fptr != pp->is_fptr
|| (*pp_found)->argc != pp->argc
|| (*pp_found)->argc_reg != pp->argc_reg
|| (*pp_found)->argc_stack != pp->argc_stack)
{ "__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 },
+ // more precise? Wine gets away with just __ftol handler
+ { "__ftol2", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
+ { "__CIpow", OPP_CIPOW, OPF_FPOP, mxST0|mxST1, mxST0 },
};
const struct parsed_proto *pp_c;
struct parsed_proto *pp;
struct parsed_data *pd;
struct parsed_op *po;
const char *tmpname;
+ enum op_op prev_op;
int i, l;
int ret;
if (pp != NULL) {
if (pp->is_fptr)
check_func_pp(po, pp, "fptr var call");
- if (pp->is_noreturn)
+ if (pp->is_noreturn) {
po->flags |= OPF_TAIL;
+ po->flags &= ~OPF_ATAIL; // most likely...
+ }
}
po->pp = pp;
continue;
tailcall:
po->op = OP_CALL;
po->flags |= OPF_TAIL;
- if (i > 0 && ops[i - 1].op == OP_POP)
+ prev_op = i > 0 ? ops[i - 1].op : OP_UD2;
+ if (prev_op == OP_POP)
+ po->flags |= OPF_ATAIL;
+ if (g_stack_fsz + g_bp_frame == 0 && prev_op != OP_PUSH
+ && (g_func_pp == NULL || g_func_pp->argc_stack > 0))
+ {
po->flags |= OPF_ATAIL;
+ }
i--; // reprocess
}
}
-static void scan_prologue_epilogue(int opcnt)
+static void scan_prologue_epilogue(int opcnt, int *stack_align)
{
int ecx_push = 0, esp_sub = 0, pusha = 0;
int sandard_epilogue;
i++;
}
+ if (ops[i].op == OP_AND && ops[i].operand[0].reg == xSP
+ && ops[i].operand[1].type == OPT_CONST)
+ {
+ l = ops[i].operand[1].val;
+ j = ffs(l) - 1;
+ if (j == -1 || (l >> j) != -1)
+ ferr(&ops[i], "unhandled esp align: %x\n", l);
+ if (stack_align != NULL)
+ *stack_align = 1 << j;
+ ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ 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;
l += ops[j].operand[1].val / 4 - 1;
}
else
- ferr(&ops[j], "'pop ecx' expected\n");
+ break;
ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
j--;
}
- if (l != ecx_push)
+ if (l != ecx_push) {
+ if (i < opcnt && ops[i].op == OP_CALL
+ && ops[i].pp != NULL && ops[i].pp->is_noreturn)
+ {
+ // noreturn tailcall with no epilogue
+ i++;
+ found = 1;
+ continue;
+ }
ferr(&ops[j], "epilogue scan failed\n");
+ }
found = 1;
}
if (esp_sub) {
if (ops[j].op != OP_ADD
|| !IS(opr_name(&ops[j], 0), "esp")
- || ops[j].operand[1].type != OPT_CONST
- || ops[j].operand[1].val != g_stack_fsz)
+ || ops[j].operand[1].type != OPT_CONST)
{
- if (ops[i].op == OP_CALL && ops[i].pp != NULL
- && ops[i].pp->is_noreturn)
+ if (i < opcnt && ops[i].op == OP_CALL
+ && ops[i].pp != NULL && ops[i].pp->is_noreturn)
{
// noreturn tailcall with no epilogue
i++;
+ found = 1;
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
+ if (ops[j].operand[1].val < g_stack_fsz)
+ ferr(&ops[j], "esp adj is too low (need %d)\n", g_stack_fsz);
+
+ ops[j].operand[1].val -= g_stack_fsz; // for stack arg scanner
+ if (ops[j].operand[1].val == 0)
+ ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
found = 1;
}
ferr(po, "too many args for '%s'\n", tmpname);
}
if (pp->argc_stack > adj / 4) {
+ if (pp->is_noreturn)
+ // assume no stack adjust was emited
+ goto out;
fnote(po, "(this call)\n");
ferr(&ops[ret], "stack tracking failed for '%s': %x %x\n",
tmpname, pp->argc_stack * 4, adj);
ferr(po, "missing esp_adjust for vararg func '%s'\n",
pp->name);
+out:
return pp;
}
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);
+ if (offset < 0 || offset >= pp->argc_stack * 4 || (offset & 3)) {
+ //ferr(po, "offset %d, %d args\n", offset, pp->argc_stack);
+ continue;
+ }
arg = base_arg + offset / 4;
po->p_argnext = -1;
return 0;
}
+static int sync_argnum(struct parsed_op *po, int argnum)
+{
+ struct parsed_op *po_tmp;
+
+ // 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;
+ }
+
+ // 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;
+ }
+
+ return argnum;
+}
+
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, int *regmask, int *arg_grp,
+ int arg, int argnum, int magic, int need_op_saving, int may_reuse)
{
struct parsed_proto *pp_tmp;
struct parsed_op *po_tmp;
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];
return -1;
}
- for (; arg < pp->argc; arg++)
+ for (; arg < pp->argc; arg++, argnum++)
if (pp->arg[arg].reg == NULL)
break;
magic = (magic & 0xffffff) | (arg << 24);
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);
+ ret = collect_call_args_r(po, lr->i, pp, regmask, arg_grp,
+ arg, argnum, magic, 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, save_arg_vars,
- arg_grp, arg, magic, need_op_saving, may_reuse);
+ ret = collect_call_args_r(po, lr->i, pp, regmask, arg_grp,
+ arg, argnum, magic, need_op_saving, may_reuse);
if (ret < 0)
return ret;
}
pp_tmp = ops[j].pp;
if (pp_tmp == NULL)
- ferr(po, "arg collect hit unparsed call '%s'\n",
- ops[j].operand[0].name);
+ ferr(po, "arg collect %d/%d hit unparsed call '%s'\n",
+ arg, pp->argc, 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);
ops[j].p_argnext = po_tmp - ops;
pp->arg[arg].datap = &ops[j];
+ argnum = sync_argnum(&ops[j], argnum);
+
need_to_save_current = 0;
- save_args = 0;
reg = -1;
if (ops[j].operand[0].type == OPT_REG)
reg = ops[j].operand[0].reg;
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;
- }
+ // mark this arg as one that needs operand saving
+ pp->arg[arg].is_saved = 1;
- if (save_args_seen & (1 << (ops[j].p_argnum - 1))) {
+ if (save_args_seen & (1 << (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);
+ argnum, pp->name);
}
}
else if (ops[j].p_argnum == 0)
{
ops[k].flags |= OPF_RMD | OPF_NOREGS | OPF_DONE;
ops[j].flags |= OPF_RMD | OPF_NOREGS | OPF_VAPUSH;
- save_args &= ~(1 << arg);
+ pp->arg[arg].is_saved = 0;
reg = -1;
}
else
ops[k].flags |= OPF_RMD | OPF_DONE;
ops[j].flags |= OPF_RMD;
ops[j].p_argpass = ret + 1;
- save_args &= ~(1 << arg);
+ pp->arg[arg].is_saved = 0;
reg = -1;
}
}
}
}
- *save_arg_vars |= save_args;
+ if (pp->arg[arg].is_saved) {
+ ops[j].flags &= ~OPF_RMD;
+ ops[j].p_argnum = argnum;
+ }
// tracking reg usage
if (reg >= 0)
*regmask |= 1 << reg;
arg++;
+ argnum++;
if (!pp->is_unresolved) {
// next arg
- for (; arg < pp->argc; arg++)
+ for (; arg < pp->argc; arg++, argnum++)
if (pp->arg[arg].reg == NULL)
break;
}
}
static int collect_call_args(struct parsed_op *po, int i,
- struct parsed_proto *pp, int *regmask, int *save_arg_vars,
- int magic)
+ 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 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);
+ ret = collect_call_args_r(po, i, pp, regmask, &arg_grp,
+ 0, 1, magic, 0, 0);
if (ret < 0)
return ret;
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;
+ po_tmp = po_tmp->p_argnext >= 0 ? &ops[po_tmp->p_argnext] : NULL;
}
}
}
- save_arg_vars[arg_grp] |= save_arg_vars_current;
if (pp->is_unresolved) {
pp->argc += ret;
pp->argc_reg++;
}
-static void output_std_flags(FILE *fout, struct parsed_op *po,
+static void output_std_flag_z(FILE *fout, struct parsed_op *po,
int *pfomask, const char *dst_opr_text)
{
if (*pfomask & (1 << PFO_Z)) {
lmod_cast_u(po, po->operand[0].lmod), dst_opr_text);
*pfomask &= ~(1 << PFO_Z);
}
+}
+
+static void output_std_flag_s(FILE *fout, struct parsed_op *po,
+ int *pfomask, const char *dst_opr_text)
+{
if (*pfomask & (1 << PFO_S)) {
fprintf(fout, "\n cond_s = (%s%s < 0);",
lmod_cast_s(po, po->operand[0].lmod), dst_opr_text);
}
}
+static void output_std_flags(FILE *fout, struct parsed_op *po,
+ int *pfomask, const char *dst_opr_text)
+{
+ output_std_flag_z(fout, po, pfomask, dst_opr_text);
+ output_std_flag_s(fout, po, pfomask, dst_opr_text);
+}
+
enum {
OPP_FORCE_NORETURN = (1 << 0),
OPP_SIMPLE_ARGS = (1 << 1),
if (!pp->is_fptr)
fprintf(fout, " a%d", i + 1);
}
+
+ if (pp->arg[i].type.is_64bit)
+ i++;
}
if (pp->is_vararg) {
if (i > 0)
int had_decl = 0;
int label_pending = 0;
int need_double = 0;
+ int stack_align = 0;
+ int stack_fsz_adj = 0;
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
// pass2:
// - handle ebp/esp frame, remove ops related to it
- scan_prologue_epilogue(opcnt);
+ scan_prologue_epilogue(opcnt, &stack_align);
+
+ // handle a case where sf size is unalignment, but is
+ // placed in a way that elements are still aligned
+ if (g_stack_fsz & 4) {
+ for (i = 0; i < g_eqcnt; i++) {
+ if (g_eqs[i].lmod != OPLM_QWORD)
+ continue;
+ if (!(g_eqs[i].offset & 4)) {
+ g_stack_fsz += 4;
+ stack_fsz_adj = 4;
+ }
+ break;
+ }
+ }
// pass3:
// - remove dead labels
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);
+ collect_call_args(po, i, pp, ®mask, i + opcnt * 2);
}
// for unresolved, collect after other passes
}
reg_use_pass(0, opcnt, cbits, regmask_init, ®mask,
0, ®mask_save, ®mask_init, regmask_arg);
+ need_float_stack = !!(regmask & mxST7_2);
+
// pass7:
// - find flag set ops for their users
// - do unresolved calls
if (pp->is_unresolved) {
int regmask_stack = 0;
- collect_call_args(po, i, pp, ®mask, save_arg_vars,
- i + opcnt * 2);
+ collect_call_args(po, i, 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)
+ if (pp->arg[arg].reg != NULL && !pp->arg[arg].is_saved)
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)
+ if (tmp_op->operand[0].type == OPT_REG)
regmask_stack |= 1 << tmp_op->operand[0].reg;
}
default:
break;
}
+
+ // this might need it's own pass...
+ if (po->op != OP_FST && po->p_argnum > 0)
+ save_arg_vars[po->p_arggrp] |= 1 << (po->p_argnum - 1);
+
+ // correct for "full stack" mode late enable
+ if ((po->flags & (OPF_PPUSH|OPF_FPOP)) && need_float_stack)
+ po->flags |= OPF_FSHIFT;
}
float_type = need_double ? "double" : "float";
- need_float_stack = !!(regmask & mxST7_2);
float_st0 = need_float_stack ? "f_st[f_stp & 7]" : "f_st0";
float_st1 = need_float_stack ? "f_st[(f_stp + 1) & 7]" : "f_st1";
// declare stack frame, va_arg
if (g_stack_fsz) {
+ if (stack_fsz_adj)
+ fprintf(fout, " // stack_fsz_adj %d\n", stack_fsz_adj);
+
fprintf(fout, " union { u32 d[%d];", (g_stack_fsz + 3) / 4);
if (g_func_lmods & (1 << OPLM_WORD))
fprintf(fout, " u16 w[%d];", (g_stack_fsz + 1) / 2);
fprintf(fout, " u8 b[%d];", g_stack_fsz);
if (g_func_lmods & (1 << OPLM_QWORD))
fprintf(fout, " double q[%d];", (g_stack_fsz + 7) / 8);
+
+ if (stack_align > 8)
+ ferr(ops, "unhandled stack align of %d\n", stack_align);
+ else if (stack_align == 8)
+ fprintf(fout, " u64 align;");
fprintf(fout, " } sf;\n");
had_decl = 1;
}
break;
case OP_BSF:
+ case OP_BSR:
+ // on SKL, if src is 0, dst is left unchanged
assert_operand_cnt(2);
+ out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
- fprintf(fout, " %s = %s ? __builtin_ffs(%s) - 1 : 0;",
- out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
- buf2, buf2);
- output_std_flags(fout, po, &pfomask, buf1);
+ output_std_flag_z(fout, po, &pfomask, buf2);
+ if (po->op == OP_BSF)
+ snprintf(buf3, sizeof(buf3), "__builtin_ffs(%s) - 1", buf2);
+ else
+ snprintf(buf3, sizeof(buf3), "31 - __builtin_clz(%s)", buf2);
+ fprintf(fout, " if (%s) %s = %s;", buf2, buf1, buf3);
last_arith_dst = &po->operand[0];
delayed_flag_op = NULL;
- strcat(g_comment, " bsf");
+ strcat(g_comment, po->op == OP_BSF ? " bsf" : " bsr");
break;
case OP_DEC:
lmod_cast_s(po, po->operand[0].lmod), buf2);
last_arith_dst = &po->operand[0];
delayed_flag_op = NULL;
- if (pfomask & (1 << PFO_C)) {
+ if (pfomask & PFOB_C) {
fprintf(fout, "\n cond_c = (%s != 0);", buf1);
- pfomask &= ~(1 << PFO_C);
+ pfomask &= ~PFOB_C;
}
+ output_std_flags(fout, po, &pfomask, buf1);
break;
case OP_IMUL:
pp->has_structarg ? "_sa" : "");
if (po->flags & OPF_ATAIL) {
- if (pp->argc_stack != g_func_pp->argc_stack
- || (pp->argc_stack > 0
- && pp->is_stdcall != g_func_pp->is_stdcall))
- ferr(po, "incompatible tailcall\n");
+ int check_compat =
+ g_func_pp->is_stdcall && g_func_pp->argc_stack > 0;
+ check_compat |= pp->argc_stack > 0;
+ if (check_compat
+ && (pp->argc_stack != g_func_pp->argc_stack
+ || pp->is_stdcall != g_func_pp->is_stdcall))
+ ferr(po, "incompatible arg-reuse tailcall\n");
if (g_func_pp->has_retreg)
ferr(po, "TODO: retreg+tailcall\n");
if (pp->arg[arg].reg != NULL) {
if (pp->arg[arg].type.is_retreg)
fprintf(fout, "&%s", pp->arg[arg].reg);
+ else if (IS(pp->arg[arg].reg, "ebp")
+ && g_bp_frame && !(po->flags & OPF_EBP_S))
+ {
+ // rare special case
+ fprintf(fout, "%s(u32)&sf.b[sizeof(sf)]", cast);
+ strcat(g_comment, " bp_ref");
+ }
else
fprintf(fout, "%s%s", cast, pp->arg[arg].reg);
continue;
else if (tmp_op->p_argpass != 0) {
fprintf(fout, "a%d", tmp_op->p_argpass);
}
- else if (tmp_op->p_argnum != 0) {
+ else if (pp->arg[arg].is_saved) {
+ ferr_assert(po, tmp_op->p_argnum > 0);
fprintf(fout, "%s%s", cast,
saved_arg_name(buf1, sizeof(buf1),
tmp_op->p_arggrp, tmp_op->p_argnum));
case OPP_ALLSHL:
case OPP_ALLSHR:
fprintf(fout, " tmp64 = ((u64)edx << 32) | eax;\n");
- fprintf(fout, " tmp64 = (s64)tmp64 %s= LOBYTE(ecx);\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
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 -> <
+ if (mask == 0x0100 || mask == 0x0500) { // C0 -> <
fprintf(fout, " f_sw = %s < %s ? 0x0100 : 0;",
float_st0, buf1);
}
fprintf(fout, " f_st0 = f_st1 * log2%s(f_st0);",
need_double ? "" : "f");
}
+ strcat(g_comment, " fyl2x");
break;
case OP_FSIN:
strcat(g_comment, " ftol");
break;
+ case OPP_CIPOW:
+ if (need_float_stack) {
+ fprintf(fout, " %s = pow%s(%s, %s);", float_st1,
+ need_double ? "" : "f", float_st1, float_st0);
+ fprintf(fout, " f_stp++;");
+ }
+ else {
+ fprintf(fout, " f_st0 = pow%s(f_st1, f_st0);",
+ need_double ? "" : "f");
+ }
+ strcat(g_comment, " CIpow");
+ break;
+
+ case OPP_ABORT:
+ fprintf(fout, " do_skip_code_abort();");
+ break;
+
// mmx
case OP_EMMS:
- strcpy(g_comment, " (emms)");
+ fprintf(fout, " do_emms();");
break;
default:
last_arith_dst = NULL;
}
- label_pending = 0;
+ if (!no_output)
+ label_pending = 0;
}
if (g_stack_fsz && !g_stack_frame_used)
po->regmask_dst |= 1 << xAX;
dep = hg_fp_find_dep(fp, po->operand[0].name);
- if (dep != NULL)
+ if (dep != NULL) {
dep->regmask_live = regmask_save | regmask_dst;
+ if (g_bp_frame && !(po->flags & OPF_EBP_S))
+ dep->regmask_live |= 1 << xBP;
+ }
}
else if (po->op == OP_RET) {
if (po->operand_cnt > 0) {
static void gen_hdr(const char *funcn, int opcnt)
{
- int save_arg_vars[MAX_ARG_GRP] = { 0, };
unsigned char cbits[MAX_OPS / 8];
const struct parsed_proto *pp_c;
struct parsed_proto *pp;
// pass2:
// - handle ebp/esp frame, remove ops related to it
- scan_prologue_epilogue(opcnt);
+ scan_prologue_epilogue(opcnt, NULL);
// pass3:
// - remove dead labels
if (!pp->is_unresolved && !(po->flags & OPF_ATAIL)) {
// since we know the args, collect them
- ret = collect_call_args(po, i, pp, ®mask_dummy, save_arg_vars,
- i + opcnt * 1);
+ ret = collect_call_args(po, i, pp, ®mask_dummy,
+ i + opcnt * 1);
}
}
}
// noreturn OS functions
break;
}
- if (ops[i].op != OP_NOP)
+ if (ops[i].op != OP_NOP && ops[i].op != OPP_ABORT)
ferr(&ops[i], "unreachable code\n");
}
long pos;
p = strrchr(p, ';');
- if (p != NULL && *p == ';' && IS_START(p + 2, "DATA XREF: ")) {
- p += 13;
- if (is_xref_needed(p, rlist, rlist_len))
+ if (p != NULL && *p == ';') {
+ if (IS_START(p + 2, "sctref"))
return 1;
+ if (IS_START(p + 2, "DATA XREF: ")) {
+ p += 13;
+ if (is_xref_needed(p, rlist, rlist_len))
+ return 1;
+ }
}
pos = ftell(fasm);
p = strrchr(p, ';');
p += 2;
+
+ if (IS_START(p, "sctref")) {
+ found_need = 1;
+ break;
+ }
+
// it's printed once, but no harm to check again
if (IS_START(p, "DATA XREF: "))
p += 11;
return strcmp(c1->name, c2->name);
}
-static void scan_ahead(FILE *fasm)
+static void scan_ahead_for_chunks(FILE *fasm)
{
char words[2][256];
char line[256];
char *sctproto = NULL;
int in_func = 0;
int pending_endp = 0;
- int skip_func = 0;
+ int skip_code = 0;
+ int skip_code_end = 0;
int skip_warned = 0;
int eq_alloc;
int verbose = 0;
verbose = 1;
else if (IS(argv[arg], "-rf"))
g_allow_regfunc = 1;
+ else if (IS(argv[arg], "-uc"))
+ g_allow_user_icall = 1;
else if (IS(argv[arg], "-m"))
multi_seg = 1;
else if (IS(argv[arg], "-hdr"))
"options:\n"
" -hdr - header generation mode\n"
" -rf - allow unannotated indirect calls\n"
+ " -uc - allow ind. calls/refs to __usercall\n"
" -m - allow multiple .text sections\n"
"[rlist] is a file with function names to skip,"
" one per line\n",
memset(words, 0, sizeof(words));
for (; arg < argc; arg++) {
+ int skip_func = 0;
+
frlist = fopen(argv[arg], "r");
my_assert_not(frlist, NULL);
}
rlist[rlist_len++] = strdup(words[0]);
}
- skip_func = 0;
fclose(frlist);
frlist = NULL;
if (addr > f_addr && !scanned_ahead) {
//anote("scan_ahead caused by '%s', addr %lx\n",
// g_func, addr);
- scan_ahead(fasm);
+ scan_ahead_for_chunks(fasm);
scanned_ahead = 1;
func_chunks_sorted = 0;
}
if (*p != 0 && *p != ';')
aerr("too many words\n");
- // alow asm patches in comments
+ if (skip_code_end) {
+ skip_code_end = 0;
+ skip_code = 0;
+ }
+
+ // allow asm patches in comments
if (*p == ';') {
if (IS_START(p, "; sctpatch:")) {
p = sskip(p + 11);
if (!pending_endp)
break;
}
+ else if (IS_START(p, "; sctskip_start")) {
+ if (in_func && !g_skip_func) {
+ if (!skip_code) {
+ ops[pi].op = OPP_ABORT;
+ ops[pi].asmln = asmln;
+ pi++;
+ }
+ skip_code = 1;
+ }
+ }
+ else if (IS_START(p, "; sctskip_end")) {
+ if (skip_code)
+ skip_code_end = 1;
+ }
}
if (wordc == 0) {
if (pd->type == OPT_OFFSET)
pd->d[pd->count].u.label = strdup(words[i]);
else
- pd->d[pd->count].u.val = parse_number(words[i]);
+ pd->d[pd->count].u.val = parse_number(words[i], 0);
pd->d[pd->count].bt_i = -1;
pd->count++;
}
if (!IS(g_func, words[0]))
aerr("endp '%s' while in_func '%s'?\n",
words[0], g_func);
+ if (skip_code)
+ aerr("endp '%s' while skipping code\n", words[0]);
if ((g_ida_func_attr & IDAFA_THUNK) && pi == 1
&& ops[0].op == OP_JMP && ops[0].operand[0].had_ds)
continue;
}
- if (!in_func || g_skip_func) {
+ if (!in_func || g_skip_func || skip_code) {
if (!skip_warned && !g_skip_func && g_labels[pi] != NULL) {
if (verbose)
anote("skipping from '%s'\n", g_labels[pi]);
else
aerr("bad lmod: '%s'\n", words[2]);
- g_eqs[g_eqcnt].offset = parse_number(words[4]);
+ g_eqs[g_eqcnt].offset = parse_number(words[4], 0);
g_eqcnt++;
continue;
}
notaz.gp2x.de / ia32rtools.git / blobdiff