def account_stack_ops(exec_file, mem_instr_addr_l):
cfg = disassm.get_func_disassm(exec_file, mem_instr_addr_l[0])
stack_op_count = 0
for pc in mem_instr_addr_l:
if pc in cfg.ins_tags_dict.keys():
if "Stack" in cfg.ins_tags_dict[pc]:
stack_op_count += 1
elif pc != 0:
cfg = disassm.get_func_disassm(exec_file, pc)
print stack_op_count, len(mem_instr_addr_l)
stack_op_ratio = float(stack_op_count)/float(len(mem_instr_addr_l))
print "Stack operation ratio: %lf"%(stack_op_ratio)
def record_time_to_update(delinq_load_addr, update_addr, trace_q, cfg, time_to_update_dict, delinq_loads_till_update, BBs_in_loop, delinq_loads_update_addr, prefetch_decisions, conf):
if not delinq_load_addr in trace_q:
return
fwd_score = 0
fwd_delinq_loads = 0
BB_addr = static_BB_cfg.discover_BB_for_address(update_addr, cfg.BB_dict)
while trace_q:
if not BB_addr in BBs_in_loop:
BBs_in_loop.append(BB_addr)
pc_in_trace = trace_q.popleft()
if pc_in_trace == 0:
return
if pc_in_trace == delinq_load_addr:
break
if pc_in_trace in cfg.ins_tags_dict:
if pc_in_trace in conf.all_delinq_loads_list and prefetch_decisions:
if prefetch_decisions[pc_in_trace].l3_mr >= 0.01:
fwd_delinq_loads += 1
elif prefetch_decisions[pc_in_trace].l2_mr >= 0.05:
fwd_delinq_loads += 1
elif prefetch_decisions[pc_in_trace].l1_mr >= 0.2:
fwd_delinq_loads += 1
fwd_score += 1
BB_addr = static_BB_cfg.discover_BB_for_address(pc_in_trace, cfg.BB_dict)
if BB_addr == None:
cfg = disassm.get_func_disassm(cfg.exec_file, pc_in_trace)
BB_addr = static_BB_cfg.discover_BB_for_address(pc_in_trace, cfg.BB_dict)
if BB_addr == None:
return
record_update_time(delinq_load_addr, fwd_score, time_to_update_dict, fwd_delinq_loads, delinq_loads_till_update)
def pointer_analysis_with_trace_hints(track_reg, is_track_idx_reg, delinq_load_addr, BB_addr, trace_q, cfg, pointer_update_addr_dict, pointer_update_time_dict, time_to_update_dict, delinq_loads_till_update, delinq_loads_till_use, delinq_loads_update_addr, update_trace_dict, prefetch_decisions, conf):
BBs_in_loop = []
update_trace_q = Queue.Queue(10)
precomp_q = Queue.Queue(11)
last_trace_pc = trace_q.popleft()
if last_trace_pc != delinq_load_addr:
return
inter_delinq_loads = 0
score = 0
is_reg_pushed_on_stack = False
equal_reg = None
while trace_q:
pc_in_trace = trace_q.popleft()
if pc_in_trace == 0:
return
BB_addr = static_BB_cfg.discover_BB_for_address(pc_in_trace, cfg.BB_dict)
if BB_addr == None:
cfg = disassm.get_func_disassm(cfg.exec_file, pc_in_trace)
BB_addr = static_BB_cfg.discover_BB_for_address(pc_in_trace, cfg.BB_dict)
if BB_addr == None:
return
if not BB_addr in BBs_in_loop:
BBs_in_loop.append(BB_addr)
tag = None
if pc_in_trace in cfg.ins_tags_dict:
tag = cfg.ins_tags_dict[pc_in_trace]
if pc_in_trace in conf.all_delinq_loads_list and prefetch_decisions:
if prefetch_decisions[pc_in_trace].l3_mr >= 0.01:
inter_delinq_loads += 1
elif prefetch_decisions[pc_in_trace].l2_mr >= 0.05:
inter_delinq_loads += 1
elif prefetch_decisions[pc_in_trace].l1_mr >= 0.2:
inter_delinq_loads += 1
score += 1
if update_trace_q.full():
print >> sys.stderr, "Too deep object nesting encountered: NestTrace @ \n"
print >> sys.stderr, update_trace_q
return BBs_in_loop
#ignore writes
if not pc_in_trace in cfg.ins_dst_regs_dict.keys():
continue
reg_updated_curr_pc = cfg.ins_dst_regs_dict[pc_in_trace][0]
# if pc_in_trace in update_trace_q.queue and tag != "Move":
if reg_updated_curr_pc == track_reg and (is_track_idx_reg or ( any(pc_in_trace in x for x in update_trace_q.queue) and tag != "Move" )):
if not pc_in_trace in pointer_update_addr_dict:
pointer_update_addr_dict[pc_in_trace] = 1
else:
pointer_update_addr_dict[pc_in_trace] += 1
# record_time_to_update(delinq_load_addr, pc_in_trace, trace_q, cfg, time_to_update_dict, delinq_loads_till_update, BBs_in_loop, delinq_loads_update_addr, prefetch_decisions, conf)
# print >> sys.stderr, ''.join('0x%02x ' % b for b in update_trace_q.queue )
# print >> sys.stderr, "%lx"%(pc_in_trace)
precomp_q.put(delinq_load_addr)
up_score = 0
if not is_track_idx_reg:
while any(pc_in_trace in x for x in update_trace_q.queue): #pc_in_trace in update_trace_q.queue:
# while update_trace_q.queue:
pc_score_t = update_trace_q.get()
pc = pc_score_t[0]
up_score = pc_score_t[1]
precomp_q.put(pc)
elif is_track_idx_reg:
pc_score_t = (pc_in_trace, score)
update_trace_q.put(pc_score_t)
while not update_trace_q.empty():
pc_score_t = update_trace_q.get()
pc = pc_score_t[0]
precomp_q.put(pc)
# precomp_q.put(pc_in_trace)
# print >> sys.stderr, ''.join('0x%02x ' % b for b in precomp_q.queue )
# while update_trace_q.queue:
# pc_score_t = update_trace_q.get()
# pc = pc_score_t[0]
# print >> sys.stderr,"%lx"%(pc)
# print >> sys.stderr, "processed precomp queue!"
record_update_time(delinq_load_addr, up_score, pointer_update_time_dict, inter_delinq_loads, delinq_loads_till_use)
record_update_addr(delinq_load_addr, pc_in_trace, delinq_loads_update_addr)
# while update_trace_q.queue:
# loop_score = update_trace_q.get()[1]
loop_score = score
fwd_score = loop_score - up_score
record_update_time(delinq_load_addr, fwd_score, time_to_update_dict, 0, delinq_loads_till_update)
precomp_q_t = tuple(list(precomp_q.queue))
record_update_trace(precomp_q_t, update_trace_dict)
return BBs_in_loop
if reg_updated_curr_pc == track_reg and not is_reg_pushed_on_stack: #the register value should not be on the stack during this
if tag == "Read":
reg_read = cfg.ins_src_regs_dict[pc_in_trace][0]
track_reg = reg_read
pc_score_t = (pc_in_trace, score)
update_trace_q.put(pc_score_t)
if cfg.ins_idx_reg_dict[pc_in_trace] != 0:
track_reg = cfg.ins_idx_reg_dict[pc_in_trace]
is_track_idx_reg = True
#print >> sys.stderr, "Tracking index register %s from pc @ %lx"%(cfg.regs_dict[track_reg], pc_in_trace)
# move r1, r2 -- not mem op
elif tag == "Move":
if pc_in_trace in cfg.ins_src_regs_dict:
reg_read = cfg.ins_src_regs_dict[pc_in_trace][0]
equal_reg = track_reg
track_reg = reg_read
else:
cfg.ins_tags_dict[pc_in_trace] = "MoveConst"
pc_score_t = (pc_in_trace, score)
update_trace_q.put(pc_score_t)
#if the register being tracked is read from the stack
elif tag == "StackR":
print >> sys.stderr,"Register pushed on stack @ %lx"%(pc_in_trace)
is_reg_pushed_on_stack = True #beyond this instruction the value is on the stack
pc_score_t = (pc_in_trace, score)
update_trace_q.put(pc_score_t)
# Not move but some instruction that changes track_reg and not traceable further (for now)
# modify this case for the LEA instruction
elif tag == "Lea":
reg_read = cfg.ins_src_regs_dict[pc_in_trace][0]
track_reg = reg_read
pc_score_t = (pc_in_trace, score)
update_trace_q.put(pc_score_t)
if cfg.ins_idx_reg_dict[pc_in_trace] != 0:
track_reg = cfg.ins_idx_reg_dict[pc_in_trace]
is_track_idx_reg = True
#print >> sys.stderr, "Tracking index register %s from pc @ %lx"%(cfg.regs_dict[track_reg], pc_in_trace)
else:
# return
# reg_read = cfg.ins_src_regs_dict[pc_in_trace][0]
# track_reg = reg_read
pc_score_t = (pc_in_trace, score)
update_trace_q.put(pc_score_t)
elif tag == "StackW":
reg_read = cfg.ins_src_regs_dict[pc_in_trace][0]
if track_reg == reg_read:
is_reg_pushed_on_stack = False #beyond this the value is in **some** register
pc_score_t = (pc_in_trace, score)
update_trace_q.put(pc_score_t)
# reaching this point is failure to find update instruction
# print >> sys.stderr, "Failed to locate update point for %lx" % (delinq_load_addr)
# print >> sys.stderr, update_trace_q
return
def pointer_analysis_with_trace_hints(track_reg, delinq_load_addr, BB_addr, trace_q, cfg, pointer_update_addr_dict, pointer_update_time_dict, time_to_update_dict, delinq_loads_till_update, delinq_loads_till_use, delinq_loads_update_addr, prefetch_decisions, conf):
# print BB_addr, BBs_inspected, pointer_update_addr_list
BBs_in_loop = []
last_trace_pc = trace_q.popleft()
if last_trace_pc != delinq_load_addr:
return
inter_delinq_loads = 0
score = 0
is_reg_pushed_on_stack = False
while trace_q:
pc_in_trace = trace_q.popleft()
if pc_in_trace == 0:
return
BB_addr = static_BB_cfg.discover_BB_for_address(pc_in_trace, cfg.BB_dict)
if BB_addr == None:
cfg = disassm.get_func_disassm(cfg.exec_file, pc_in_trace)
BB_addr = static_BB_cfg.discover_BB_for_address(pc_in_trace, cfg.BB_dict)
if BB_addr == None:
return
if not BB_addr in BBs_in_loop:
BBs_in_loop.append(BB_addr)
tag = None
if pc_in_trace in cfg.ins_tags_dict:
tag = cfg.ins_tags_dict[pc_in_trace]
if pc_in_trace in conf.all_delinq_loads_list and prefetch_decisions:
if prefetch_decisions[pc_in_trace].l3_mr >= 0.01:
inter_delinq_loads += 1
elif prefetch_decisions[pc_in_trace].l2_mr >= 0.05:
inter_delinq_loads += 1
elif prefetch_decisions[pc_in_trace].l1_mr >= 0.2:
inter_delinq_loads += 1
score += 1
if not pc_in_trace in cfg.ins_dst_regs_dict.keys():
continue
reg_updated_curr_pc = cfg.ins_dst_regs_dict[pc_in_trace][0]
if reg_updated_curr_pc == track_reg and not is_reg_pushed_on_stack: #the register value should not be on the track during this
if tag == "Read":
if not pc_in_trace in pointer_update_addr_dict:
pointer_update_addr_dict[pc_in_trace] = 1
else:
pointer_update_addr_dict[pc_in_trace] += 1
# if(is_nested_object(delinq_load_addr, pc_in_trace, cfg)):
# print >> sys.stderr, "Nested Object found @ 0x%lx --> 0x%lx!!!" % (delinq_load_addr, pc_in_trace)
# track_reg = cfg.ins_base_reg_dict[pc_in_trace]
# continue
track_reg = None
#should not include the pointer update instruction, its latency should not be counted
# score -= 1
record_update_time(delinq_load_addr, score, pointer_update_time_dict, inter_delinq_loads, delinq_loads_till_use)
record_update_addr(delinq_load_addr, pc_in_trace, delinq_loads_update_addr)
record_time_to_update(delinq_load_addr, pc_in_trace, trace_q, cfg, time_to_update_dict, delinq_loads_till_update, BBs_in_loop, delinq_loads_update_addr, prefetch_decisions, conf)
return BBs_in_loop
# return track_reg
# move r1, r2 -- not mem op
elif tag == "Move":
reg_read = cfg.ins_src_regs_dict[pc_in_trace][0]
track_reg = reg_read
#if the register being tracked is read from the stack
elif tag == "StackR":
is_reg_pushed_on_stack = True #beyond this instruction the value is on the stack
# Not move but some instruction that changes track_reg and not traceable further (for now)
# modify this case for the LEA instruction
else:
if not pc_in_trace in pointer_update_addr_dict:
pointer_update_addr_dict[pc_in_trace] = 1
else:
pointer_update_addr_dict[pc_in_trace] += 1
track_reg = None
# score -= 1
record_update_time(delinq_load_addr, score, pointer_update_time_dict, inter_delinq_loads, delinq_loads_till_use)
record_update_addr(delinq_load_addr, pc_in_trace, delinq_loads_update_addr)
record_time_to_update(delinq_load_addr, pc_in_trace, trace_q, cfg, time_to_update_dict, delinq_loads_till_update, BBs_in_loop, delinq_loads_update_addr, prefetch_decisions, conf)
return BBs_in_loop
elif tag == "StackW":
reg_read = cfg.ins_src_regs_dict[pc_in_trace][0]
if track_reg == reg_read:
is_reg_pushed_on_stack = False #beyond this the value is in **some** register
# reaching this point is failure to find update instruction
return
def analyze_non_strided_delinq_loads(global_pc_smptrace_hist, global_pc_stride_hist, prefetch_decisions, exec_file, num_samples, avg_mem_latency):
ins_src_regs_dict = {}
ins_dst_regs_dict = {}
ins_tags_dict = {}
branch_dict = {}
routine_BB_dict = {}
# information maps for Memory operations
ins_base_reg_dict = {}
ins_mem_dis_dict = {}
ins_idx_reg_dict = {}
ins_mem_scale_dict = {}
global_prefetchable_pcs = []
delinq_load_address_list = []
for delinq_load_addr in prefetch_decisions.keys():
pref_param = prefetch_decisions[delinq_load_addr]
if "ptr" in pref_param.pf_type:
delinq_load_address_list.append(delinq_load_addr)
delinq_load_address_list = sorted(delinq_load_address_list)
conf = Conf1(exec_file, delinq_load_address_list, num_samples, avg_mem_latency)
irr_list = []
print >> sys.stderr, "\nSample freq irregular accesses!\n"
for pc in delinq_load_address_list:
pc_smptrace_hist = global_pc_smptrace_hist[pc]
l3mr = prefetch_decisions[pc].l3_mr
l2mr = prefetch_decisions[pc].l2_mr
l1mr = prefetch_decisions[pc].l1_mr
sample_freq = float(len(pc_smptrace_hist.keys()))/float(num_samples)
score = float(sample_freq)*float(l3mr)
irr_list += [(pc, sample_freq, l3mr, l2mr, l1mr, score)]
sorted_irr_list = sorted(irr_list, key=operator.itemgetter(5), reverse=True)
trimmed_delinq_load_addr_list = []
count = 0
for tup in sorted_irr_list:
pc = tup[0]
sample_freq = tup[1]
l3mr = tup[2]
l2mr = tup[3]
l1mr = tup[4]
score = tup[5]
if count < 15:
trimmed_delinq_load_addr_list += [pc]
count = count + 1
pc_stride_hist = global_pc_stride_hist[pc]
sorted_x = sorted(pc_stride_hist.iteritems(), key=operator.itemgetter(1), reverse=True)
sample_count = sum([pair[1] for pair in sorted_x])
max_stride = sorted_x[0][0]
max_stride_freq = float(sorted_x[0][1])/float(sample_count)
print >> sys.stderr, "\npc:%lx freq:%lf l3mr:%lf l2mr:%lf l1mr:%lf score:%lf"%(pc, sample_freq, l3mr, l2mr, l1mr, score)
for delinq_load_addr in trimmed_delinq_load_addr_list: #delinq_load_address_list:
cfg = disassm.get_func_disassm(conf.exec_file, delinq_load_addr)
if not (cfg.ins_tags_dict[delinq_load_addr] == 'Read' or cfg.ins_tags_dict[delinq_load_addr] == 'Write'):
continue
# if float(len(global_pc_smptrace_hist[delinq_load_addr].keys()))/float(conf.num_samples) < 0.005:
# continue
(pointer_update_addr_dict, pointer_update_time_dict, time_to_update_dict, delinq_loads_till_update, delinq_loads_till_use, all_BBs_in_loop, is_ind, stride) = ins_trace_ptr_nobj_analysis.detect_pointer_chasing(global_pc_smptrace_hist, global_pc_stride_hist, delinq_load_addr, None, cfg, conf)
analyze_pointer_prefetch(pointer_update_addr_dict, prefetch_decisions, pointer_update_time_dict, time_to_update_dict, delinq_load_addr, delinq_loads_till_update, delinq_loads_till_use, all_BBs_in_loop, cfg, conf, is_ind, stride)
# (pointer_update_addr_dict, pointer_update_time_dict, time_to_update_dict, delinq_loads_till_update, delinq_loads_till_use, all_BBs_in_loop) = ins_trace_analysis.detect_pointer_chasing(global_pc_smptrace_hist, delinq_load_addr, prefetch_decisions, cfg, conf)
# analyze_pointer_prefetch(pointer_update_addr_dict, pointer_update_time_dict, time_to_update_dict, delinq_load_addr, delinq_loads_till_update, delinq_loads_till_use, all_BBs_in_loop, cfg, conf)
if delinq_load_addr in conf.indirect_pref_decisions:
do_cost_benefit_analysis(cfg, conf, delinq_load_addr, prefetch_decisions)
decide_prefetch_schedules(cfg, conf)
print_indirect_prefetch_decisions(conf)
except IOError, e:
continue
usf_file.close()
for (pc_rdist_hist, pc_stride_hist, pc_freq_hist, pc_time_hist, pc_corr_hist, pc_fwd_rdist_hist, pc_smptrace_hist) in burst_hists:
continue
ins_trace_ptr_nobj_analysis.add_trace_to_global_pc_smptrace_hist(global_pc_smptrace_hist, pc_smptrace_hist)
ins_trace_ptr_nobj_analysis.add_to_pc_stride_hist(pc_stride_hist, global_pc_stride_hist)
print >> sys.stderr, "Starting trace analysis..."
for delinq_load_addr in delinq_load_address_list:
cfg = disassm.get_func_disassm(conf.exec_file, delinq_load_addr)
if not (cfg.ins_tags_dict[delinq_load_addr] == 'Read' or cfg.ins_tags_dict[delinq_load_addr] == 'Write'):
continue
print >> sys.stderr, "Sample frequency %lx: %lf"%(delinq_load_addr, float(len(pc_smptrace_hist.keys()))/float(conf.num_samples))
(pointer_update_addr_dict, pointer_update_time_dict, time_to_update_dict, delinq_loads_till_update, delinq_loads_till_use, all_BBs_in_loop, is_ind, stride) = ins_trace_ptr_nobj_analysis.detect_pointer_chasing(global_pc_smptrace_hist, global_pc_stride_hist, delinq_load_addr, None, cfg, conf)
# analyze_pointer_prefetch(pointer_update_addr_dict, pointer_update_time_dict, time_to_update_dict, delinq_load_addr, delinq_loads_till_update, delinq_loads_till_use, all_BBs_in_loop, cfg, conf, is_ind, stride)
analyze_pointer_prefetch(pointer_update_addr_dict, [], pointer_update_time_dict, time_to_update_dict, delinq_load_addr, delinq_loads_till_update, delinq_loads_till_use, all_BBs_in_loop, cfg, conf, is_ind, stride)
decide_prefetch_schedules(cfg, conf)
print_indirect_prefetch_decisions(conf)
