/* returns false on failure */
static bool
decode_function(void *dcontext, byte *entry)
{
byte *pc, *pre_pc;
int num_instr = 0;
bool found_ret = false;
instr_t *instr;
if (entry == NULL)
return false;
instr = instr_create(dcontext);
pc = entry;
while (true) {
instr_reset(dcontext, instr);
pre_pc = pc;
pc = decode(dcontext, pc, instr);
instr_set_translation(instr, pre_pc);
dr_print_instr(dcontext, STDOUT, instr, "");
if (instr_is_return(instr)) {
found_ret = true;
break;
}
num_instr++;
if (num_instr > MAX_INSTRS_IN_FUNCTION) {
print("ERROR: hit max instr limit %d\n", MAX_INSTRS_IN_FUNCTION);
break;
}
}
instr_destroy(dcontext, instr);
return found_ret;
}
static void
look_for_usercall(void *dcontext, byte *entry, const char *sym, LOADED_IMAGE *img,
const char *modpath)
{
bool found_push_imm = false;
int imm = 0;
app_pc pc, pre_pc;
instr_t *instr;
if (entry == NULL)
return;
instr = instr_create(dcontext);
pc = entry;
while (true) {
instr_reset(dcontext, instr);
pre_pc = pc;
pc = decode(dcontext, pc, instr);
if (verbose) {
instr_set_translation(instr, pre_pc);
dr_print_instr(dcontext, STDOUT, instr, "");
}
if (pc == NULL || !instr_valid(instr))
break;
if (instr_get_opcode(instr) == OP_push_imm) {
found_push_imm = true;
imm = (int) opnd_get_immed_int(instr_get_src(instr, 0));
} else if (instr_is_call_direct(instr) && found_push_imm) {
app_pc tgt = opnd_get_pc(instr_get_target(instr));
bool found = false;
int i;
for (i = 0; i < NUM_USERCALL; i++) {
if (tgt == usercall_addr[i]) {
dr_printf("Call #0x%02x to %s at %s+0x%x\n", imm, usercall_names[i],
sym, pre_pc - entry);
found = true;
break;
}
}
if (found)
break;
} else if (instr_is_return(instr))
break;
if (pc - entry > MAX_BYTES_BEFORE_USERCALL)
break;
}
instr_destroy(dcontext, instr);
}
/* This event is passed the instruction list for the whole bb. */
static dr_emit_flags_t
event_analyze_bb(void *drcontext, void *tag, instrlist_t *bb,
bool for_trace, bool translating, void **user_data)
{
/* Count the instructions and pass the result to event_insert_instrumentation. */
per_bb_data_t *per_bb = dr_thread_alloc(drcontext, sizeof(*per_bb));
instr_t *instr;
uint num_instrs = 0;
uint num_flops = 0;
uint num_syscalls = 0;
dr_fp_type_t fp_type;
for (instr = instrlist_first_app(bb);
instr != NULL;
instr = instr_get_next_app(instr)) {
num_instrs++;
if (instr_is_floating_ex(instr, &fp_type) &&
/* We exclude loads and stores (and reg-reg moves) and state preservation */
(fp_type == DR_FP_CONVERT || fp_type == DR_FP_MATH)) {
#ifdef VERBOSE
dr_print_instr(drcontext, STDOUT, instr, "Found flop: ");
#endif
num_flops++;
}
if (instr_is_syscall(instr)) {
num_syscalls++;
}
}
per_bb->num_instrs = num_instrs;
per_bb->num_flops = num_flops;
per_bb->num_syscalls = num_syscalls;
*(per_bb_data_t**)user_data = per_bb;
return DR_EMIT_DEFAULT;
}
std::string
raw2trace_t::append_bb_entries(uint tidx, offline_entry_t *in_entry, OUT bool *handled)
{
uint instr_count = in_entry->pc.instr_count;
instr_t instr;
trace_entry_t buf_start[MAX_COMBINED_ENTRIES];
app_pc start_pc = modvec[in_entry->pc.modidx].map_base + in_entry->pc.modoffs;
app_pc pc, decode_pc = start_pc;
if ((in_entry->pc.modidx == 0 && in_entry->pc.modoffs == 0) ||
modvec[in_entry->pc.modidx].map_base == NULL) {
// FIXME i#2062: add support for code not in a module (vsyscall, JIT, etc.).
// Once that support is in we can remove the bool return value and handle
// the memrefs up here.
VPRINT(3, "Skipping ifetch for %u instrs not in a module\n", instr_count);
*handled = false;
return "";
} else {
VPRINT(3, "Appending %u instrs in bb " PFX " in mod %u +" PIFX " = %s\n",
instr_count, (ptr_uint_t)start_pc, (uint)in_entry->pc.modidx,
(ptr_uint_t)in_entry->pc.modoffs, modvec[in_entry->pc.modidx].path);
}
bool skip_icache = false;
if (instr_count == 0) {
// L0 filtering adds a PC entry with a count of 0 prior to each memref.
skip_icache = true;
instr_count = 1;
// We set a flag to avoid peeking forward on instr entries.
if (!instrs_are_separate)
instrs_are_separate = true;
}
CHECK(!instrs_are_separate || instr_count == 1, "cannot mix 0-count and >1-count");
instr_init(dcontext, &instr);
for (uint i = 0; i < instr_count; ++i) {
trace_entry_t *buf = buf_start;
app_pc orig_pc = decode_pc - modvec[in_entry->pc.modidx].map_base +
modvec[in_entry->pc.modidx].orig_base;
bool skip_instr = false;
instr_reset(dcontext, &instr);
// We assume the default ISA mode and currently require the 32-bit
// postprocessor for 32-bit applications.
pc = decode(dcontext, decode_pc, &instr);
if (pc == NULL || !instr_valid(&instr)) {
WARN("Encountered invalid/undecodable instr @ %s+" PFX,
modvec[in_entry->pc.modidx].path, (ptr_uint_t)in_entry->pc.modoffs);
break;
}
CHECK(!instr_is_cti(&instr) || i == instr_count - 1, "invalid cti");
if (instr_is_rep_string(&instr)) {
// We want it to look like the original rep string instead of the
// drutil-expanded loop.
if (!prev_instr_was_rep_string)
prev_instr_was_rep_string = true;
else
skip_instr = true;
} else
prev_instr_was_rep_string = false;
// FIXME i#1729: make bundles via lazy accum until hit memref/end.
if (!skip_instr) {
DO_VERBOSE(3, {
instr_set_translation(&instr, orig_pc);
dr_print_instr(dcontext, STDOUT, &instr, "");
});
/* replaces inc with add 1, dec with sub 1
* returns true if successful, false if not
*/
static bool
replace_inc_with_add(void *drcontext, instr_t *instr, instrlist_t *trace)
{
instr_t *in;
uint eflags;
int opcode = instr_get_opcode(instr);
bool ok_to_replace = false;
DR_ASSERT(opcode == OP_inc || opcode == OP_dec);
#ifdef VERBOSE
dr_print_instr(drcontext, STDOUT, instr, "in replace_inc_with_add:\n\t");
#endif
/* add/sub writes CF, inc/dec does not, make sure that's ok */
for (in = instr; in != NULL; in = instr_get_next(in)) {
eflags = instr_get_eflags(in);
if ((eflags & EFLAGS_READ_CF) != 0) {
#ifdef VERBOSE
dr_print_instr(drcontext, STDOUT, in,
"\treads CF => cannot replace inc with add: ");
#endif
return false;
}
if (instr_is_exit_cti(in)) {
/* to be more sophisticated, examine instructions at
* target of exit cti (if it is a direct branch).
* for this example, we give up if we hit a branch.
*/
return false;
}
/* if writes but doesn't read, ok */
if ((eflags & EFLAGS_WRITE_CF) != 0) {
ok_to_replace = true;
break;
}
}
if (!ok_to_replace) {
#ifdef VERBOSE
dr_printf("\tno write to CF => cannot replace inc with add\n");
#endif
return false;
}
if (opcode == OP_inc) {
#ifdef VERBOSE
dr_printf("\treplacing inc with add\n");
#endif
in = INSTR_CREATE_add(drcontext, instr_get_dst(instr, 0),
OPND_CREATE_INT8(1));
} else {
#ifdef VERBOSE
dr_printf("\treplacing dec with sub\n");
#endif
in = INSTR_CREATE_sub(drcontext, instr_get_dst(instr, 0),
OPND_CREATE_INT8(1));
}
if (instr_get_prefix_flag(instr, PREFIX_LOCK))
instr_set_prefix_flag(in, PREFIX_LOCK);
instr_set_translation(in, instr_get_app_pc(instr));
instrlist_replace(trace, instr, in);
instr_destroy(drcontext, instr);
return true;
}
/* returns false on failure */
static bool
decode_syscall_num(void *dcontext, byte *entry, syscall_info_t *info, LOADED_IMAGE *img)
{
/* FIXME: would like to fail gracefully rather than have a DR assertion
* on non-code! => use DEBUG=0 INTERNAL=1 DR build!
*/
bool found_syscall = false, found_eax = false, found_edx = false, found_ecx = false;
bool found_ret = false;
byte *pc, *pre_pc;
int num_instr = 0;
instr_t *instr;
byte *preferred = get_preferred_base(img);
if (entry == NULL)
return false;
info->num_args = -1; /* if find sysnum but not args */
info->sysnum = -1;
info->fixup_index = -1;
instr = instr_create(dcontext);
pc = entry;
/* FIXME - we don't support decoding 64bit instructions in 32bit mode, but I want
* this to work on 32bit machines. Hack fix based on the wrapper pattern, we skip
* the first instruction (mov r10, rcx) here, the rest should decode ok.
* Xref PR 236203. */
if (expect_x64 && *pc == 0x4c && *(pc+1) == 0x8b && *(pc+2) == 0xd1)
pc += 3;
while (true) {
instr_reset(dcontext, instr);
pre_pc = pc;
pc = decode(dcontext, pc, instr);
if (verbose) {
instr_set_translation(instr, pre_pc);
dr_print_instr(dcontext, STDOUT, instr, "");
}
if (pc == NULL || !instr_valid(instr))
break;
if (instr_is_syscall(instr) || instr_is_call_indirect(instr)) {
/* If we see a syscall instr or an indirect call which is not syscall,
* we assume this is not a syscall wrapper.
*/
found_syscall = process_syscall_instr(dcontext, instr, found_eax, found_edx);
if (!found_syscall)
break; /* assume not a syscall wrapper, give up gracefully */
} else if (instr_is_return(instr)) {
/* we must break on return to avoid case like win8 x86
* which has sysenter callee adjacent-"inlined"
* ntdll!NtYieldExecution:
* 77d7422c b801000000 mov eax,1
* 77d74231 e801000000 call ntdll!NtYieldExecution+0xb (77d74237)
* 77d74236 c3 ret
* 77d74237 8bd4 mov edx,esp
* 77d74239 0f34 sysenter
* 77d7423b c3 ret
*/
if (!found_ret) {
process_ret(instr, info);
found_ret = true;
}
break;
} else if (instr_get_opcode(instr) == OP_call) {
found_syscall = process_syscall_call(dcontext, pc, instr,
found_eax, found_edx);
/* If we see a call and it is not a sysenter callee,
* we assume this is not a syscall wrapper.
*/
if (!found_syscall)
break; /* assume not a syscall wrapper, give up gracefully */
} else if (instr_is_cti(instr)) {
/* We expect only ctis like ret or ret imm, syscall, and call, which are
* handled above. Give up gracefully if we hit any other cti.
* XXX: what about jmp to shared ret (seen in the past on some syscalls)?
*/
/* Update: win10 TH2 1511 x64 has a cti:
* ntdll!NtContinue:
* 00007ff9`13185630 4c8bd1 mov r10,rcx
* 00007ff9`13185633 b843000000 mov eax,43h
* 00007ff9`13185638 f604250803fe7f01 test byte ptr [SharedUserData+0x308 (00000000`7ffe0308)],1
* 00007ff9`13185640 7503 jne ntdll!NtContinue+0x15 (00007ff9`13185645)
* 00007ff9`13185642 0f05 syscall
* 00007ff9`13185644 c3 ret
* 00007ff9`13185645 cd2e int 2Eh
* 00007ff9`13185647 c3 ret
*/
if (expect_x64 && instr_is_cbr(instr) &&
opnd_get_pc(instr_get_target(instr)) == pc + 3/*syscall;ret*/) {
/* keep going */
} else
break;
} else if ((!found_eax || !found_edx || !found_ecx) &&
instr_get_opcode(instr) == OP_mov_imm &&
opnd_is_reg(instr_get_dst(instr, 0))) {
if (!found_eax && opnd_get_reg(instr_get_dst(instr, 0)) == REG_EAX) {
info->sysnum = (int) opnd_get_immed_int(instr_get_src(instr, 0));
found_eax = true;
} else if (!found_edx && opnd_get_reg(instr_get_dst(instr, 0)) == REG_EDX) {
uint imm = (uint) opnd_get_immed_int(instr_get_src(instr, 0));
if (imm == 0x7ffe0300 ||
/* On Win10 the immed is ntdll!Wow64SystemServiceCall */
(expect_wow && imm > (ptr_uint_t)preferred &&
imm < (ptr_uint_t)preferred + img->SizeOfImage))
found_edx = true;
} else if (!found_ecx && opnd_get_reg(instr_get_dst(instr, 0)) == REG_ECX) {
found_ecx = true;
info->fixup_index = (int) opnd_get_immed_int(instr_get_src(instr, 0));
}
} else if (instr_get_opcode(instr) == OP_xor &&
opnd_is_reg(instr_get_src(instr, 0)) &&
opnd_get_reg(instr_get_src(instr, 0)) == REG_ECX &&
opnd_is_reg(instr_get_dst(instr, 0)) &&
opnd_get_reg(instr_get_dst(instr, 0)) == REG_ECX) {
/* xor to 0 */
found_ecx = true;
info->fixup_index = 0;
}
num_instr++;
if (num_instr > MAX_INSTRS_BEFORE_SYSCALL) /* wrappers should be short! */
break; /* avoid weird cases like NPXEMULATORTABLE */
}
instr_destroy(dcontext, instr);
return found_syscall;
}
/* returns whether found a syscall
* - found_eax: whether the caller has seen "mov imm => %eax"
* - found_edx: whether the caller has seen "mov $0x7ffe0300 => %edx",
* xref the comment in process_syscall_instr.
*/
static bool
process_syscall_call(void *dcontext, byte *next_pc, instr_t *call,
bool found_eax, bool found_edx)
{
int num_instr;
byte *pc;
instr_t instr;
bool found_syscall = false;
assert(instr_get_opcode(call) == OP_call && opnd_is_pc(instr_get_target(call)));
pc = opnd_get_pc(instr_get_target(call));
if (pc > next_pc + MAX_SYSENTER_CALLEE_OFFSET ||
pc <= next_pc /* assuming the call won't go backward */)
return false;
/* handle win8 x86 which has sysenter callee adjacent-"inlined"
* ntdll!NtYieldExecution:
* 77d7422c b801000000 mov eax,1
* 77d74231 e801000000 call ntdll!NtYieldExecution+0xb (77d74237)
* 77d74236 c3 ret
* 77d74237 8bd4 mov edx,esp
* 77d74239 0f34 sysenter
* 77d7423b c3 ret
*
* or DrMem-i#1366-c#2
* USER32!NtUserCreateWindowStation:
* 75caea7a b841110000 mov eax,0x1141
* 75caea7f e838000000 call user32!...+0xd (75caeabc)
* 75caea84 c22000 ret 0x20
* ...
* USER32!GetWindowStationName:
* 75caea8c 8bff mov edi,edi
* 75caea8e 55 push ebp
* ...
* 75caeabc 8bd4 mov edx,esp
* 75caeabe 0f34 sysenter
* 75caeac0 c3 ret
*/
/* We expect the win8 x86 sysenter adjacent "inlined" callee to be as simple as
* 75caeabc 8bd4 mov edx,esp
* 75caeabe 0f34 sysenter
* 75caeac0 c3 ret
*/
instr_init(dcontext, &instr);
num_instr = 0;
do {
instr_reset(dcontext, &instr);
pc = decode(dcontext, pc, &instr);
if (verbose)
dr_print_instr(dcontext, STDOUT, &instr, "");
if (pc == NULL || !instr_valid(&instr))
break;
if (instr_is_syscall(&instr) || instr_is_call_indirect(&instr)) {
found_syscall = process_syscall_instr(dcontext, &instr, found_eax, found_edx);
break;
} else if (instr_is_cti(&instr)) {
break;
}
num_instr++;
} while (num_instr <= MAX_INSTRS_SYSENTER_CALLEE);
instr_free(dcontext, &instr);
return found_syscall;
}
/* returns false on failure */
static bool
decode_syscall_num(void *dcontext, byte *entry, syscall_info_t *info)
{
/* FIXME: would like to fail gracefully rather than have a DR assertion
* on non-code! => use DEBUG=0 INTERNAL=1 DR build!
*/
bool found_syscall = false, found_eax = false, found_edx = false, found_ecx = false;
bool found_ret = false;
byte *pc;
int num_instr = 0;
instr_t *instr;
if (entry == NULL)
return false;
info->num_args = -1; /* if find sysnum but not args */
info->sysnum = -1;
info->fixup_index = -1;
instr = instr_create(dcontext);
pc = entry;
/* FIXME - we don't support decoding 64bit instructions in 32bit mode, but I want
* this to work on 32bit machines. Hack fix based on the wrapper pattern, we skip
* the first instruction (mov r10, rcx) here, the rest should decode ok.
* Xref PR 236203. */
if (expect_x64 && *pc == 0x4c && *(pc+1) == 0x8b && *(pc+2) == 0xd1)
pc += 3;
while (true) {
instr_reset(dcontext, instr);
pc = decode(dcontext, pc, instr);
if (verbose)
dr_print_instr(dcontext, STDOUT, instr, "");
if (pc == NULL || !instr_valid(instr))
break;
/* ASSUMPTION: a mov imm of 0x7ffe0300 into edx followed by an
* indirect call via edx is a system call on XP and later
* On XP SP1 it's call *edx, while on XP SP2 it's call *(edx)
* For wow it's a call through fs.
* FIXME - core exports various is_*_syscall routines (such as
* instr_is_wow64_syscall()) which we could use here instead of
* duplicating if they were more flexible about when they could
* be called (instr_is_wow64_syscall() for ex. asserts if not
* in a wow process).
*/
if (/* int 2e or x64 or win8 sysenter */
(instr_is_syscall(instr) && found_eax && (expect_int2e || expect_x64 || expect_sysenter)) ||
/* sysenter case */
(expect_sysenter && found_edx && found_eax &&
instr_is_call_indirect(instr) &&
/* XP SP{0,1}, 2003 SP0: call *edx */
((opnd_is_reg(instr_get_target(instr)) &&
opnd_get_reg(instr_get_target(instr)) == REG_EDX) ||
/* XP SP2, 2003 SP1: call *(edx) */
(opnd_is_base_disp(instr_get_target(instr)) &&
opnd_get_base(instr_get_target(instr)) == REG_EDX &&
opnd_get_index(instr_get_target(instr)) == REG_NULL &&
opnd_get_disp(instr_get_target(instr)) == 0))) ||
/* wow case
* we don't require found_ecx b/c win8 does not use ecx
*/
(expect_wow && found_eax &&
instr_is_call_indirect(instr) &&
opnd_is_far_base_disp(instr_get_target(instr)) &&
opnd_get_base(instr_get_target(instr)) == REG_NULL &&
opnd_get_index(instr_get_target(instr)) == REG_NULL &&
opnd_get_segment(instr_get_target(instr)) == SEG_FS)) {
found_syscall = true;
} else if (instr_is_return(instr)) {
if (!found_ret) {
process_ret(instr, info);
found_ret = true;
}
break;
} else if (instr_is_cti(instr)) {
if (instr_get_opcode(instr) == OP_call) {
/* handle win8 x86 which has sysenter callee adjacent-"inlined"
* ntdll!NtYieldExecution:
* 77d7422c b801000000 mov eax,1
* 77d74231 e801000000 call ntdll!NtYieldExecution+0xb (77d74237)
* 77d74236 c3 ret
* 77d74237 8bd4 mov edx,esp
* 77d74239 0f34 sysenter
* 77d7423b c3 ret
*/
byte *tgt;
assert(opnd_is_pc(instr_get_target(instr)));
tgt = opnd_get_pc(instr_get_target(instr));
/* we expect only ret or ret imm, and possibly some nops (in gdi32).
* XXX: what about jmp to shared ret (seen in the past on some syscalls)?
*/
if (tgt > pc && tgt <= pc + 16) {
bool ok = false;
do {
if (pc == tgt) {
ok = true;
break;
}
instr_reset(dcontext, instr);
pc = decode(dcontext, pc, instr);
if (verbose)
dr_print_instr(dcontext, STDOUT, instr, "");
if (instr_is_return(instr)) {
process_ret(instr, info);
found_ret = true;
} else if (!instr_is_nop(instr))
break;
num_instr++;
} while (num_instr <= MAX_INSTRS_BEFORE_SYSCALL);
if (ok)
continue;
}
}
/* assume not a syscall wrapper if we hit a cti */
break; /* give up gracefully */
} else if ((!found_eax || !found_edx || !found_ecx) &&
instr_get_opcode(instr) == OP_mov_imm &&
opnd_is_reg(instr_get_dst(instr, 0))) {
if (!found_eax && opnd_get_reg(instr_get_dst(instr, 0)) == REG_EAX) {
info->sysnum = (int) opnd_get_immed_int(instr_get_src(instr, 0));
found_eax = true;
} else if (!found_edx && opnd_get_reg(instr_get_dst(instr, 0)) == REG_EDX) {
int imm = (int) opnd_get_immed_int(instr_get_src(instr, 0));
if (imm == 0x7ffe0300)
found_edx = true;
} else if (!found_ecx && opnd_get_reg(instr_get_dst(instr, 0)) == REG_ECX) {
found_ecx = true;
info->fixup_index = (int) opnd_get_immed_int(instr_get_src(instr, 0));
}
} else if (instr_get_opcode(instr) == OP_xor &&
opnd_is_reg(instr_get_src(instr, 0)) &&
opnd_get_reg(instr_get_src(instr, 0)) == REG_ECX &&
opnd_is_reg(instr_get_dst(instr, 0)) &&
opnd_get_reg(instr_get_dst(instr, 0)) == REG_ECX) {
/* xor to 0 */
found_ecx = true;
info->fixup_index = 0;
}
num_instr++;
if (num_instr > MAX_INSTRS_BEFORE_SYSCALL) /* wrappers should be short! */
break; /* avoid weird cases like NPXEMULATORTABLE */
}
instr_destroy(dcontext, instr);
return found_syscall;
}
