static dr_emit_flags_t
event_bb_insert(void *drcontext, void *tag, instrlist_t *bb, instr_t *inst,
bool for_trace, bool translating, void *user_data)
{
static int freq;
reg_id_t reg1 = IF_X86_ELSE(DR_REG_XAX, DR_REG_R0);
reg_id_t reg2 = IF_X86_ELSE(DR_REG_XCX, DR_REG_R1);
CHECK(drmgr_is_first_instr(drcontext, instrlist_first_app(bb)), "first incorrect");
CHECK(!drmgr_is_first_instr(drcontext, instrlist_last(bb)) ||
instrlist_first_app(bb) == instrlist_last(bb), "first incorrect");
CHECK(drmgr_is_last_instr(drcontext, instrlist_last(bb)), "last incorrect");
CHECK(!drmgr_is_last_instr(drcontext, instrlist_first_app(bb)) ||
instrlist_first_app(bb) == instrlist_last(bb), "last incorrect");
/* hack to instrument every nth bb. assumes DR serializes bb events. */
freq++;
if (freq % 100 == 0 && inst == (instr_t*)user_data/*first instr*/) {
/* test read from cache */
dr_save_reg(drcontext, bb, inst, reg1, SPILL_SLOT_1);
drmgr_insert_read_tls_field(drcontext, tls_idx, bb, inst, reg1);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_tls_from_cache,
false, 1, opnd_create_reg(reg1));
drmgr_insert_read_cls_field(drcontext, cls_idx, bb, inst, reg1);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_cls_from_cache,
false, 1, opnd_create_reg(reg1));
dr_restore_reg(drcontext, bb, inst, reg1, SPILL_SLOT_1);
}
if (freq % 300 == 0 && inst == (instr_t*)user_data/*first instr*/) {
instr_t *first, *second;
/* test write from cache */
dr_save_reg(drcontext, bb, inst, reg1, SPILL_SLOT_1);
dr_save_reg(drcontext, bb, inst, reg2, SPILL_SLOT_2);
instrlist_insert_mov_immed_ptrsz(drcontext,
(ptr_int_t)MAGIC_NUMBER_FROM_CACHE,
opnd_create_reg(reg1),
bb, inst, &first, &second);
instr_set_meta(first);
if (second != NULL)
instr_set_meta(second);
drmgr_insert_write_tls_field(drcontext, tls_idx, bb, inst, reg1, reg2);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_tls_write_from_cache,
false, 0);
drmgr_insert_write_cls_field(drcontext, cls_idx, bb, inst, reg1, reg2);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_cls_write_from_cache,
false, 0);
dr_restore_reg(drcontext, bb, inst, reg2, SPILL_SLOT_2);
dr_restore_reg(drcontext, bb, inst, reg1, SPILL_SLOT_1);
}
return DR_EMIT_DEFAULT;
}
static void
insert_counter_update(void *drcontext, instrlist_t *bb, instr_t *where, int offset)
{
/* Since the inc instruction clobbers 5 of the arithmetic eflags,
* we have to save them around the inc. We could be more efficient
* by not bothering to save the overflow flag and constructing our
* own sequence of instructions to save the other 5 flags (using
* lahf).
*/
if (drreg_reserve_aflags(drcontext, bb, where) != DRREG_SUCCESS) {
DR_ASSERT(false); /* cannot recover */
return;
}
/* Increment the global counter using the lock prefix to make it atomic
* across threads. It would be cheaper to aggregate the thread counters
* in the exit events, but this sample is intended to illustrate inserted
* instrumentation.
*/
instrlist_meta_preinsert(
bb, where,
LOCK(INSTR_CREATE_inc(
drcontext, OPND_CREATE_ABSMEM(((byte *)&global_count) + offset, OPSZ_4))));
/* Increment the thread private counter. */
if (dr_using_all_private_caches()) {
per_thread_t *data = (per_thread_t *)drmgr_get_tls_field(drcontext, tls_idx);
/* private caches - we can use an absolute address */
instrlist_meta_preinsert(
bb, where,
INSTR_CREATE_inc(drcontext,
OPND_CREATE_ABSMEM(((byte *)&data) + offset, OPSZ_4)));
} else {
/* shared caches - we must indirect via thread local storage */
reg_id_t scratch;
if (drreg_reserve_register(drcontext, bb, where, NULL, &scratch) != DRREG_SUCCESS)
DR_ASSERT(false);
drmgr_insert_read_tls_field(drcontext, tls_idx, bb, where, scratch);
instrlist_meta_preinsert(
bb, where, INSTR_CREATE_inc(drcontext, OPND_CREATE_MEM32(scratch, offset)));
if (drreg_unreserve_register(drcontext, bb, where, scratch) != DRREG_SUCCESS)
DR_ASSERT(false);
}
if (drreg_unreserve_aflags(drcontext, bb, where) != DRREG_SUCCESS)
DR_ASSERT(false); /* cannot recover */
}
static dr_emit_flags_t
event_bb_insert(void *drcontext, void *tag, instrlist_t *bb, instr_t *inst,
bool for_trace, bool translating, void *user_data)
{
/* hack to instrument every nth bb. assumes DR serializes bb events. */
static int freq;
freq++;
if (freq % 100 == 0 && inst == (instr_t*)user_data/*first instr*/) {
/* test read from cache */
dr_save_reg(drcontext, bb, inst, DR_REG_XAX, SPILL_SLOT_1);
drmgr_insert_read_tls_field(drcontext, tls_idx, bb, inst, DR_REG_XAX);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_tls_from_cache,
false, 1, opnd_create_reg(DR_REG_XAX));
drmgr_insert_read_cls_field(drcontext, cls_idx, bb, inst, DR_REG_XAX);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_cls_from_cache,
false, 1, opnd_create_reg(DR_REG_XAX));
dr_restore_reg(drcontext, bb, inst, DR_REG_XAX, SPILL_SLOT_1);
}
if (freq % 300 == 0 && inst == (instr_t*)user_data/*first instr*/) {
/* test write from cache */
dr_save_reg(drcontext, bb, inst, DR_REG_XAX, SPILL_SLOT_1);
dr_save_reg(drcontext, bb, inst, DR_REG_XCX, SPILL_SLOT_2);
instrlist_meta_preinsert(bb, inst, INSTR_CREATE_mov_imm
(drcontext, opnd_create_reg(DR_REG_EAX),
OPND_CREATE_INT32(MAGIC_NUMBER_FROM_CACHE)));
drmgr_insert_write_tls_field(drcontext, tls_idx, bb, inst, DR_REG_XAX,
DR_REG_XCX);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_tls_write_from_cache,
false, 0);
drmgr_insert_write_cls_field(drcontext, cls_idx, bb, inst, DR_REG_XAX,
DR_REG_XCX);
dr_insert_clean_call(drcontext, bb, inst, (void *)check_cls_write_from_cache,
false, 0);
dr_restore_reg(drcontext, bb, inst, DR_REG_XCX, SPILL_SLOT_2);
dr_restore_reg(drcontext, bb, inst, DR_REG_XAX, SPILL_SLOT_1);
}
return DR_EMIT_DEFAULT;
}
static void dynamic_info_instrumentation(void *drcontext, instrlist_t *ilist, instr_t *where,
instr_t * static_info)
{
/*
issues that may arise
1. pc and eflags is uint but in 64 bit mode 8 byte transfers are done -> so far no problem (need to see this)
need to see whether there is a better way
2. double check all the printing
*/
/*
this function does the acutal instrumentation
arguments -
we get a filled pointer here about the operand types for a given instruction (srcs and dests)
1) increment the pointer to the instr_trace buffers
2) add this pointer to instr_trace_t wrapper
3) check whether any of the srcs and dests have memory operations; if so add a lea instruction and get the dynamic address
Add this address to instr_trace_t structure
4) if the buffer is full call a function to dump it to the file and restore the head ptr of the buffer
(lean function is used utilizing a code cache to limit code bloat needed for a clean call before every instruction.)
*/
instr_t *instr, *call, *restore, *first, *second;
opnd_t ref, opnd1, opnd2;
reg_id_t reg1 = DR_REG_XBX; /* We can optimize it by picking dead reg */
reg_id_t reg2 = DR_REG_XCX; /* reg2 must be ECX or RCX for jecxz */
reg_id_t reg3 = DR_REG_XAX;
per_thread_t *data;
uint pc;
uint i;
module_data_t * module_data;
if (client_arg->instrace_mode == DISASSEMBLY_TRACE){
dr_insert_clean_call(drcontext, ilist, where, clean_call_disassembly_trace, false, 0);
return;
}
data = drmgr_get_tls_field(drcontext, tls_index);
/* Steal the register for memory reference address *
* We can optimize away the unnecessary register save and restore
* by analyzing the code and finding the register is dead.
*/
dr_save_reg(drcontext, ilist, where, reg1, SPILL_SLOT_2);
dr_save_reg(drcontext, ilist, where, reg2, SPILL_SLOT_3);
dr_save_reg(drcontext, ilist, where, reg3, SPILL_SLOT_4);
drmgr_insert_read_tls_field(drcontext, tls_index, ilist, where, reg2);
/* Load data->buf_ptr into reg2 */
opnd1 = opnd_create_reg(reg2);
opnd2 = OPND_CREATE_MEMPTR(reg2, offsetof(per_thread_t, buf_ptr));
instr = INSTR_CREATE_mov_ld(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* buf_ptr->static_info_instr = static_info; */
/* Move static_info to static_info_instr field of buf (which is a instr_trace_t *) */
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(instr_trace_t, static_info_instr));
instrlist_insert_mov_immed_ptrsz(drcontext, (ptr_int_t)static_info, opnd1, ilist, where, &first, &second);
/* buf_ptr->num_mem = 0; */
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(instr_trace_t, num_mem));
instrlist_insert_mov_immed_ptrsz(drcontext, (ptr_int_t)0, opnd1, ilist, where, &first, &second);
for (i = 0; i<instr_num_dsts(where); i++){
if (opnd_is_memory_reference(instr_get_dst(where, i))){
ref = instr_get_dst(where, i);
DR_ASSERT(opnd_is_null(ref) == false);
dr_restore_reg(drcontext, ilist, where, reg1, SPILL_SLOT_2);
dr_restore_reg(drcontext, ilist, where, reg2, SPILL_SLOT_3);
#ifdef DEBUG_MEM_REGS
dr_insert_clean_call(drcontext, ilist, where, clean_call_disassembly_trace, false, 0);
dr_insert_clean_call(drcontext, ilist, where, clean_call_print_regvalues, false, 0);
#endif
drutil_insert_get_mem_addr(drcontext, ilist, where, ref, reg1, reg2);
#ifdef DEBUG_MEM_REGS
dr_insert_clean_call(drcontext, ilist, where, clean_call_print_regvalues, false, 0);
#endif
#ifdef DEBUG_MEM_STATS
dr_insert_clean_call(drcontext, ilist, where, clean_call_disassembly_trace, false, 0);
dr_insert_clean_call(drcontext, ilist, where, clean_call_mem_stats, false, 1, opnd_create_reg(reg1));
#endif
dr_insert_clean_call(drcontext, ilist, where, clean_call_populate_mem, false, 3, opnd_create_reg(reg1), OPND_CREATE_INT32(i), OPND_CREATE_INT32(DST_TYPE));
}
}
for (i = 0; i<instr_num_srcs(where); i++){
if (opnd_is_memory_reference(instr_get_src(where, i))){
ref = instr_get_src(where, i);
DR_ASSERT(opnd_is_null(ref) == false);
dr_restore_reg(drcontext, ilist, where, reg1, SPILL_SLOT_2);
dr_restore_reg(drcontext, ilist, where, reg2, SPILL_SLOT_3);
#ifdef DEBUG_MEM_REGS
dr_insert_clean_call(drcontext, ilist, where, clean_call_disassembly_trace, false, 0);
dr_insert_clean_call(drcontext, ilist, where, clean_call_print_regvalues, false, 0);
#endif
drutil_insert_get_mem_addr(drcontext, ilist, where, ref, reg1, reg2);
#ifdef DEBUG_MEM_REGS
dr_insert_clean_call(drcontext, ilist, where, clean_call_print_regvalues, false, 0);
#endif
#ifdef DEBUG_MEM_STATS
dr_insert_clean_call(drcontext, ilist, where, clean_call_disassembly_trace, false, 0);
dr_insert_clean_call(drcontext, ilist, where, clean_call_mem_stats, false, 1, opnd_create_reg(reg1));
#endif
dr_insert_clean_call(drcontext, ilist, where, clean_call_populate_mem, false, 3, opnd_create_reg(reg1), OPND_CREATE_INT32(i), OPND_CREATE_INT32(SRC_TYPE));
}
}
drmgr_insert_read_tls_field(drcontext, tls_index, ilist, where, reg2);
/* Load data->buf_ptr into reg2 */
opnd1 = opnd_create_reg(reg2);
opnd2 = OPND_CREATE_MEMPTR(reg2, offsetof(per_thread_t, buf_ptr));
instr = INSTR_CREATE_mov_ld(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* arithmetic flags are saved here for buf_ptr->eflags filling */
dr_save_arith_flags_to_xax(drcontext, ilist, where);
/* load the eflags */
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(instr_trace_t, eflags));
opnd2 = opnd_create_reg(reg3);
instr = INSTR_CREATE_mov_st(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* load the app_pc */
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(instr_trace_t, pc));
module_data = dr_lookup_module(instr_get_app_pc(where));
//dynamically generated code - module information not available - then just store 0 at the pc slot of the instr_trace data
if (module_data != NULL){
pc = instr_get_app_pc(where) - module_data->start;
dr_free_module_data(module_data);
}
else{
pc = 0;
}
instrlist_insert_mov_immed_ptrsz(drcontext, (ptr_int_t)pc, opnd1, ilist, where, &first, &second);
/* buf_ptr++; */
/* Increment reg value by pointer size using lea instr */
opnd1 = opnd_create_reg(reg2);
opnd2 = opnd_create_base_disp(reg2, DR_REG_NULL, 0,
sizeof(instr_trace_t),
OPSZ_lea);
instr = INSTR_CREATE_lea(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Update the data->buf_ptr */
drmgr_insert_read_tls_field(drcontext, tls_index, ilist, where, reg1);
opnd1 = OPND_CREATE_MEMPTR(reg1, offsetof(per_thread_t, buf_ptr));
opnd2 = opnd_create_reg(reg2);
instr = INSTR_CREATE_mov_st(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* we use lea + jecxz trick for better performance
* lea and jecxz won't disturb the eflags, so we won't insert
* code to save and restore application's eflags.
*/
/* lea [reg2 - buf_end] => reg2 */
opnd1 = opnd_create_reg(reg1);
opnd2 = OPND_CREATE_MEMPTR(reg1, offsetof(per_thread_t, buf_end));
instr = INSTR_CREATE_mov_ld(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
opnd1 = opnd_create_reg(reg2);
opnd2 = opnd_create_base_disp(reg1, reg2, 1, 0, OPSZ_lea);
instr = INSTR_CREATE_lea(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* jecxz call */
call = INSTR_CREATE_label(drcontext);
opnd1 = opnd_create_instr(call);
instr = INSTR_CREATE_jecxz(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* jump restore to skip clean call */
restore = INSTR_CREATE_label(drcontext);
opnd1 = opnd_create_instr(restore);
instr = INSTR_CREATE_jmp(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* clean call */
/* We jump to lean procedure which performs full context switch and
* clean call invocation. This is to reduce the code cache size.
*/
instrlist_meta_preinsert(ilist, where, call);
/* mov restore DR_REG_XCX */
opnd1 = opnd_create_reg(reg2);
/* this is the return address for jumping back from lean procedure */
opnd2 = opnd_create_instr(restore);
/* We could use instrlist_insert_mov_instr_addr(), but with a register
* destination we know we can use a 64-bit immediate.
*/
instr = INSTR_CREATE_mov_imm(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* jmp code_cache */
opnd1 = opnd_create_pc(code_cache);
instr = INSTR_CREATE_jmp(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* restore %reg */
instrlist_meta_preinsert(ilist, where, restore);
//dr_restore_arith_flags_from_xax(drcontext, ilist, where);
dr_restore_reg(drcontext, ilist, where, reg1, SPILL_SLOT_2);
dr_restore_reg(drcontext, ilist, where, reg2, SPILL_SLOT_3);
dr_restore_reg(drcontext, ilist, where, reg3, SPILL_SLOT_4);
//instrlist_disassemble(drcontext, instr_get_app_pc(instrlist_first(ilist)), ilist, logfile);
}
/* instrument_instr is called whenever a memory reference is identified.
* It inserts code before the memory reference to to fill the memory buffer
* and jump to our own code cache to call the clean_call when the buffer is full.
*/
static void
instrument_instr(void *drcontext, instrlist_t *ilist, instr_t *where)
{
instr_t *instr, *call, *restore;
opnd_t opnd1, opnd2;
reg_id_t reg1, reg2;
drvector_t allowed;
per_thread_t *data;
app_pc pc;
data = drmgr_get_tls_field(drcontext, tls_index);
/* Steal two scratch registers.
* reg2 must be ECX or RCX for jecxz.
*/
drreg_init_and_fill_vector(&allowed, false);
drreg_set_vector_entry(&allowed, DR_REG_XCX, true);
if (drreg_reserve_register(drcontext, ilist, where, &allowed, ®2) !=
DRREG_SUCCESS ||
drreg_reserve_register(drcontext, ilist, where, NULL, ®1) != DRREG_SUCCESS) {
DR_ASSERT(false); /* cannot recover */
drvector_delete(&allowed);
return;
}
drvector_delete(&allowed);
/* The following assembly performs the following instructions
* buf_ptr->pc = pc;
* buf_ptr->opcode = opcode;
* buf_ptr++;
* if (buf_ptr >= buf_end_ptr)
* clean_call();
*/
drmgr_insert_read_tls_field(drcontext, tls_index, ilist, where, reg2);
/* Load data->buf_ptr into reg2 */
opnd1 = opnd_create_reg(reg2);
opnd2 = OPND_CREATE_MEMPTR(reg2, offsetof(per_thread_t, buf_ptr));
instr = INSTR_CREATE_mov_ld(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Store pc */
pc = instr_get_app_pc(where);
/* For 64-bit, we can't use a 64-bit immediate so we split pc into two halves.
* We could alternatively load it into reg1 and then store reg1.
* We use a convenience routine that does the two-step store for us.
*/
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(ins_ref_t, pc));
instrlist_insert_mov_immed_ptrsz(drcontext, (ptr_int_t) pc, opnd1,
ilist, where, NULL, NULL);
/* Store opcode */
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(ins_ref_t, opcode));
opnd2 = OPND_CREATE_INT32(instr_get_opcode(where));
instr = INSTR_CREATE_mov_st(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Increment reg value by pointer size using lea instr */
opnd1 = opnd_create_reg(reg2);
opnd2 = opnd_create_base_disp(reg2, DR_REG_NULL, 0, sizeof(ins_ref_t), OPSZ_lea);
instr = INSTR_CREATE_lea(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Update the data->buf_ptr */
drmgr_insert_read_tls_field(drcontext, tls_index, ilist, where, reg1);
opnd1 = OPND_CREATE_MEMPTR(reg1, offsetof(per_thread_t, buf_ptr));
opnd2 = opnd_create_reg(reg2);
instr = INSTR_CREATE_mov_st(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* We use the lea + jecxz trick for better performance.
* lea and jecxz won't disturb the eflags, so we won't need
* code to save and restore the application's eflags.
*/
/* lea [reg2 - buf_end] => reg2 */
opnd1 = opnd_create_reg(reg1);
opnd2 = OPND_CREATE_MEMPTR(reg1, offsetof(per_thread_t, buf_end));
instr = INSTR_CREATE_mov_ld(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
opnd1 = opnd_create_reg(reg2);
opnd2 = opnd_create_base_disp(reg1, reg2, 1, 0, OPSZ_lea);
instr = INSTR_CREATE_lea(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* jecxz call */
call = INSTR_CREATE_label(drcontext);
opnd1 = opnd_create_instr(call);
instr = INSTR_CREATE_jecxz(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* jump restore to skip clean call */
restore = INSTR_CREATE_label(drcontext);
opnd1 = opnd_create_instr(restore);
instr = INSTR_CREATE_jmp(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* clean call */
/* We jump to our generated lean procedure which performs a full context
* switch and clean call invocation. This is to reduce the code cache size.
*/
instrlist_meta_preinsert(ilist, where, call);
/* mov restore DR_REG_XCX */
opnd1 = opnd_create_reg(reg2);
/* This is the return address for jumping back from the lean procedure. */
opnd2 = opnd_create_instr(restore);
/* We could use instrlist_insert_mov_instr_addr(), but with a register
* destination we know we can use a 64-bit immediate.
*/
instr = INSTR_CREATE_mov_imm(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* jmp code_cache */
opnd1 = opnd_create_pc(code_cache);
instr = INSTR_CREATE_jmp(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* Restore scratch registers */
instrlist_meta_preinsert(ilist, where, restore);
if (drreg_unreserve_register(drcontext, ilist, where, reg1) != DRREG_SUCCESS ||
drreg_unreserve_register(drcontext, ilist, where, reg2) != DRREG_SUCCESS)
DR_ASSERT(false);
}
/*
* instrument_mem is called whenever a memory reference is identified.
* It inserts code before the memory reference to to fill the memory buffer
* and jump to our own code cache to call the clean_call when the buffer is full.
*/
static void
instrument_mem(void *drcontext, instrlist_t *ilist, instr_t *where,
int pos, bool write)
{
instr_t *instr, *call, *restore, *first, *second;
opnd_t ref, opnd1, opnd2;
reg_id_t reg1 = DR_REG_XBX; /* We can optimize it by picking dead reg */
reg_id_t reg2 = DR_REG_XCX; /* reg2 must be ECX or RCX for jecxz */
per_thread_t *data;
app_pc pc;
data = drmgr_get_tls_field(drcontext, tls_index);
/* Steal the register for memory reference address *
* We can optimize away the unnecessary register save and restore
* by analyzing the code and finding the register is dead.
*/
dr_save_reg(drcontext, ilist, where, reg1, SPILL_SLOT_2);
dr_save_reg(drcontext, ilist, where, reg2, SPILL_SLOT_3);
if (write)
ref = instr_get_dst(where, pos);
else
ref = instr_get_src(where, pos);
/* use drutil to get mem address */
drutil_insert_get_mem_addr(drcontext, ilist, where, ref, reg1, reg2);
/* The following assembly performs the following instructions
* buf_ptr->write = write;
* buf_ptr->addr = addr;
* buf_ptr->size = size;
* buf_ptr->pc = pc;
* buf_ptr++;
* if (buf_ptr >= buf_end_ptr)
* clean_call();
*/
drmgr_insert_read_tls_field(drcontext, tls_index, ilist, where, reg2);
/* Load data->buf_ptr into reg2 */
opnd1 = opnd_create_reg(reg2);
opnd2 = OPND_CREATE_MEMPTR(reg2, offsetof(per_thread_t, buf_ptr));
instr = INSTR_CREATE_mov_ld(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Move write/read to write field */
opnd1 = OPND_CREATE_MEM32(reg2, offsetof(mem_ref_t, write));
opnd2 = OPND_CREATE_INT32(write);
instr = INSTR_CREATE_mov_imm(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Store address in memory ref */
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(mem_ref_t, addr));
opnd2 = opnd_create_reg(reg1);
instr = INSTR_CREATE_mov_st(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Store size in memory ref */
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(mem_ref_t, size));
/* drutil_opnd_mem_size_in_bytes handles OP_enter */
opnd2 = OPND_CREATE_INT32(drutil_opnd_mem_size_in_bytes(ref, where));
instr = INSTR_CREATE_mov_st(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Store pc in memory ref */
pc = instr_get_app_pc(where);
/* For 64-bit, we can't use a 64-bit immediate so we split pc into two halves.
* We could alternatively load it into reg1 and then store reg1.
* We use a convenience routine that does the two-step store for us.
*/
opnd1 = OPND_CREATE_MEMPTR(reg2, offsetof(mem_ref_t, pc));
instrlist_insert_mov_immed_ptrsz(drcontext, (ptr_int_t) pc, opnd1,
ilist, where, &first, &second);
instr_set_ok_to_mangle(first, false/*meta*/);
if (second != NULL)
instr_set_ok_to_mangle(second, false/*meta*/);
/* Increment reg value by pointer size using lea instr */
opnd1 = opnd_create_reg(reg2);
opnd2 = opnd_create_base_disp(reg2, DR_REG_NULL, 0,
sizeof(mem_ref_t),
OPSZ_lea);
instr = INSTR_CREATE_lea(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* Update the data->buf_ptr */
drmgr_insert_read_tls_field(drcontext, tls_index, ilist, where, reg1);
opnd1 = OPND_CREATE_MEMPTR(reg1, offsetof(per_thread_t, buf_ptr));
opnd2 = opnd_create_reg(reg2);
instr = INSTR_CREATE_mov_st(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* we use lea + jecxz trick for better performance
* lea and jecxz won't disturb the eflags, so we won't insert
* code to save and restore application's eflags.
*/
/* lea [reg2 - buf_end] => reg2 */
opnd1 = opnd_create_reg(reg1);
opnd2 = OPND_CREATE_MEMPTR(reg1, offsetof(per_thread_t, buf_end));
instr = INSTR_CREATE_mov_ld(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
opnd1 = opnd_create_reg(reg2);
opnd2 = opnd_create_base_disp(reg1, reg2, 1, 0, OPSZ_lea);
instr = INSTR_CREATE_lea(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* jecxz call */
call = INSTR_CREATE_label(drcontext);
opnd1 = opnd_create_instr(call);
instr = INSTR_CREATE_jecxz(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* jump restore to skip clean call */
restore = INSTR_CREATE_label(drcontext);
opnd1 = opnd_create_instr(restore);
instr = INSTR_CREATE_jmp(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* clean call */
/* We jump to lean procedure which performs full context switch and
* clean call invocation. This is to reduce the code cache size.
*/
instrlist_meta_preinsert(ilist, where, call);
/* mov restore DR_REG_XCX */
opnd1 = opnd_create_reg(reg2);
/* this is the return address for jumping back from lean procedure */
opnd2 = opnd_create_instr(restore);
/* We could use instrlist_insert_mov_instr_addr(), but with a register
* destination we know we can use a 64-bit immediate.
*/
instr = INSTR_CREATE_mov_imm(drcontext, opnd1, opnd2);
instrlist_meta_preinsert(ilist, where, instr);
/* jmp code_cache */
opnd1 = opnd_create_pc(code_cache);
instr = INSTR_CREATE_jmp(drcontext, opnd1);
instrlist_meta_preinsert(ilist, where, instr);
/* restore %reg */
instrlist_meta_preinsert(ilist, where, restore);
dr_restore_reg(drcontext, ilist, where, reg1, SPILL_SLOT_2);
dr_restore_reg(drcontext, ilist, where, reg2, SPILL_SLOT_3);
}
