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); }
/* PR 215143: auto-magically add size prefixes */ static void test_size_changes(void *dc) { /* * 0x004299d4 67 51 addr16 push %ecx %sp -> %sp (%sp) * 0x004299d4 66 51 data16 push %cx %esp -> %esp (%esp) * 0x004299d4 66 67 51 data16 addr16 push %cx %sp -> %sp (%sp) * 0x004298a4 e3 fe jecxz $0x004298a4 %ecx * 0x004298a4 67 e3 fd addr16 jecxz $0x004298a4 %cx * 0x080a5260 67 e2 fd addr16 loop $0x080a5260 %cx -> %cx * 0x080a5260 67 e1 fd addr16 loope $0x080a5260 %cx -> %cx * 0x080a5260 67 e0 fd addr16 loopne $0x080a5260 %cx -> %cx */ instr_t *instr; /* push addr16 */ instr = instr_create_2dst_2src(dc, OP_push, opnd_create_reg(IF_X64_ELSE(REG_ESP, REG_SP)), opnd_create_base_disp(IF_X64_ELSE(REG_ESP, REG_SP), REG_NULL, 0, -(int)sizeof(void*), OPSZ_ret), opnd_create_reg(REG_XCX), opnd_create_reg(IF_X64_ELSE(REG_ESP, REG_SP))); test_instr_encode(dc, instr, 2); #ifndef X64 /* can only shorten on AMD */ /* push data16 */ instr = instr_create_2dst_2src(dc, OP_push, opnd_create_reg(REG_XSP), opnd_create_base_disp(REG_XSP, REG_NULL, 0, -2, OPSZ_2), opnd_create_reg(REG_CX), opnd_create_reg(REG_XSP)); test_instr_encode(dc, instr, 2); /* push addr16 and data16 */ instr = instr_create_2dst_2src(dc, OP_push, opnd_create_reg(REG_SP), opnd_create_base_disp(REG_SP, REG_NULL, 0, -2, OPSZ_2), opnd_create_reg(REG_CX), opnd_create_reg(REG_SP)); test_instr_encode(dc, instr, 3); #endif /* jecxz and jcxz */ test_instr_encode(dc, INSTR_CREATE_jecxz(dc, opnd_create_pc(buf)), 2); /* test non-default count register size (requires addr prefix) */ instr = instr_create_0dst_2src (dc, OP_jecxz, opnd_create_pc(buf), opnd_create_reg(IF_X64_ELSE(REG_ECX, REG_CX))); test_instr_encode(dc, instr, 3); instr = instr_create_1dst_2src (dc, OP_loop, opnd_create_reg(IF_X64_ELSE(REG_ECX, REG_CX)), opnd_create_pc(buf), opnd_create_reg(IF_X64_ELSE(REG_ECX, REG_CX))); test_instr_encode(dc, instr, 3); instr = instr_create_1dst_2src (dc, OP_loope, opnd_create_reg(IF_X64_ELSE(REG_ECX, REG_CX)), opnd_create_pc(buf), opnd_create_reg(IF_X64_ELSE(REG_ECX, REG_CX))); test_instr_encode(dc, instr, 3); instr = instr_create_1dst_2src (dc, OP_loopne, opnd_create_reg(IF_X64_ELSE(REG_ECX, REG_CX)), opnd_create_pc(buf), opnd_create_reg(IF_X64_ELSE(REG_ECX, REG_CX))); test_instr_encode(dc, instr, 3); /* * 0x004ee0b8 a6 cmps %ds:(%esi) %es:(%edi) %esi %edi -> %esi %edi * 0x004ee0b8 67 a6 addr16 cmps %ds:(%si) %es:(%di) %si %di -> %si %di * 0x004ee0b8 66 a7 data16 cmps %ds:(%esi) %es:(%edi) %esi %edi -> %esi %edi * 0x004ee0b8 d7 xlat %ds:(%ebx,%al,1) -> %al * 0x004ee0b8 67 d7 addr16 xlat %ds:(%bx,%al,1) -> %al * 0x004ee0b8 0f f7 c1 maskmovq %mm0 %mm1 -> %ds:(%edi) * 0x004ee0b8 67 0f f7 c1 addr16 maskmovq %mm0 %mm1 -> %ds:(%di) * 0x004ee0b8 66 0f f7 c1 maskmovdqu %xmm0 %xmm1 -> %ds:(%edi) * 0x004ee0b8 67 66 0f f7 c1 addr16 maskmovdqu %xmm0 %xmm1 -> %ds:(%di) */ test_instr_encode(dc, INSTR_CREATE_cmps_1(dc), 1); instr = instr_create_2dst_4src (dc, OP_cmps, opnd_create_reg(IF_X64_ELSE(REG_ESI, REG_SI)), opnd_create_reg(IF_X64_ELSE(REG_EDI, REG_DI)), opnd_create_far_base_disp(SEG_DS, IF_X64_ELSE(REG_ESI, REG_SI), REG_NULL, 0, 0, OPSZ_1), opnd_create_far_base_disp(SEG_ES, IF_X64_ELSE(REG_EDI, REG_DI), REG_NULL, 0, 0, OPSZ_1), opnd_create_reg(IF_X64_ELSE(REG_ESI, REG_SI)), opnd_create_reg(IF_X64_ELSE(REG_EDI, REG_DI))); test_instr_encode(dc, instr, 2); instr = instr_create_2dst_4src (dc, OP_cmps, opnd_create_reg(REG_XSI), opnd_create_reg(REG_XDI), opnd_create_far_base_disp(SEG_DS, REG_XSI, REG_NULL, 0, 0, OPSZ_2), opnd_create_far_base_disp(SEG_ES, REG_XDI, REG_NULL, 0, 0, OPSZ_2), opnd_create_reg(REG_XSI), opnd_create_reg(REG_XDI)); test_instr_encode_and_decode(dc, instr, 2, true/*src*/, 0, OPSZ_2, 2); test_instr_encode(dc, INSTR_CREATE_xlat(dc), 1); instr = instr_create_1dst_1src (dc, OP_xlat, opnd_create_reg(REG_AL), opnd_create_far_base_disp(SEG_DS, IF_X64_ELSE(REG_EBX, REG_BX), REG_AL, 1, 0, OPSZ_1)); test_instr_encode(dc, instr, 2); instr = INSTR_CREATE_maskmovq(dc, opnd_create_reg(REG_MM0), opnd_create_reg(REG_MM1)); test_instr_encode(dc, instr, 3); instr = instr_create_1dst_2src (dc, OP_maskmovq, opnd_create_far_base_disp(SEG_DS, IF_X64_ELSE(REG_EDI, REG_DI), REG_NULL, 0, 0, OPSZ_8), opnd_create_reg(REG_MM0), opnd_create_reg(REG_MM1)); test_instr_encode(dc, instr, 4); instr = INSTR_CREATE_maskmovdqu(dc, opnd_create_reg(REG_XMM0), opnd_create_reg(REG_XMM1)); test_instr_encode(dc, instr, 4); instr = instr_create_1dst_2src (dc, OP_maskmovdqu, opnd_create_far_base_disp(SEG_DS, IF_X64_ELSE(REG_EDI, REG_DI), REG_NULL, 0, 0, OPSZ_16), opnd_create_reg(REG_XMM0), opnd_create_reg(REG_XMM1)); test_instr_encode(dc, instr, 5); /* Test iretw, iretd, iretq (unlike most stack operation iretd (and lretd on AMD) * exist and are the default in 64-bit mode. As such, it has a different size/type * then most other stack operations). Our instr_create routine should match stack * (iretq on 64-bit, iretd on 32-bit). See PR 191977. */ instr = INSTR_CREATE_iret(dc); #ifdef X64 test_instr_encode_and_decode(dc, instr, 2, true /*src*/, 1, OPSZ_40, 40); ASSERT(buf[0] == 0x48); /* check for rex.w prefix */ #else test_instr_encode_and_decode(dc, instr, 1, true /*src*/, 1, OPSZ_12, 12); #endif instr = instr_create_1dst_2src (dc, OP_iret, opnd_create_reg(REG_XSP), opnd_create_reg(REG_XSP), opnd_create_base_disp(REG_XSP, REG_NULL, 0, 0, OPSZ_12)); test_instr_encode_and_decode(dc, instr, 1, true /*src*/, 1, OPSZ_12, 12); instr = instr_create_1dst_2src (dc, OP_iret, opnd_create_reg(REG_XSP), opnd_create_reg(REG_XSP), opnd_create_base_disp(REG_XSP, REG_NULL, 0, 0, OPSZ_6)); test_instr_encode_and_decode(dc, instr, 2, true /*src*/, 1, OPSZ_6, 6); ASSERT(buf[0] == 0x66); /* check for data prefix */ }
/* * 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); }