/* Our version, versus Dr. Memory's version in drmemory/fastpath.c */ bool instr_ok_for_instrument_fastpath(instr_t *inst, fastpath_info_t *mi, bb_info_t *bi) { uint opc = instr_get_opcode(inst); int i; initialize_fastpath_info(mi, bi, inst); if (!options.fastpath) return false; if (opc == OP_xlat) { /* can't use base-disp "%ds:(%ebx,%al,1)" for lea: would have to expand * to multiple instrs. not worth supporting since pretty rare though I * do see 3K in twolf test on windows. */ return false; } /* We assume that any one memory reference, even in a rep string form, * will only access one heap allocation. Since we're taking the most * recent access to any part of a heap alloc we thus don't care about the * size of a memory reference. */ for (i=0; i<instr_num_dsts(inst); i++) { if (opnd_uses_memory_we_track(instr_get_dst(inst, i))) { mi->store = true; if (!opnd_is_null(mi->dst[0].app)) { /* FIXME: we could handle 2 dsts if no srcs easily, * and even more dsts if we really wanted to w/o too * much trouble. also something like pusha w/ * consecutive dsts is easy: just take first one. */ return false; } mi->dst[0].app = instr_get_dst(inst, i); } } for (i=0; i<instr_num_srcs(inst); i++) { if (opnd_uses_memory_we_track(instr_get_src(inst, i))) { if (mi->store) mi->mem2mem = true; else mi->load = true; if (!opnd_is_null(mi->src[0].app)) { /* see notes above about handling this: in particular cmps */ return false; } mi->src[0].app = instr_get_src(inst, i); } } return true; }
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); }