cpython/Tools/scripts/analyze_dxp.py

131 lines
4.1 KiB
Python
Executable File

#!/usr/bin/env python
"""
Some helper functions to analyze the output of sys.getdxp() (which is
only available if Python was built with -DDYNAMIC_EXECUTION_PROFILE).
These will tell you which opcodes have been executed most frequently
in the current process, and, if Python was also built with -DDXPAIRS,
will tell you which instruction _pairs_ were executed most frequently,
which may help in choosing new instructions.
If Python was built without -DDYNAMIC_EXECUTION_PROFILE, importing
this module will raise a RuntimeError.
If you're running a script you want to profile, a simple way to get
the common pairs is:
$ PYTHONPATH=$PYTHONPATH:<python_srcdir>/Tools/scripts \
./python -i -O the_script.py --args
...
> from analyze_dxp import *
> s = render_common_pairs()
> open('/tmp/some_file', 'w').write(s)
"""
import copy
import opcode
import operator
import sys
import threading
if not hasattr(sys, "getdxp"):
raise RuntimeError("Can't import analyze_dxp: Python built without"
" -DDYNAMIC_EXECUTION_PROFILE.")
_profile_lock = threading.RLock()
_cumulative_profile = sys.getdxp()
# If Python was built with -DDXPAIRS, sys.getdxp() returns a list of
# lists of ints. Otherwise it returns just a list of ints.
def has_pairs(profile):
"""Returns True if the Python that produced the argument profile
was built with -DDXPAIRS."""
return len(profile) > 0 and isinstance(profile[0], list)
def reset_profile():
"""Forgets any execution profile that has been gathered so far."""
with _profile_lock:
sys.getdxp() # Resets the internal profile
global _cumulative_profile
_cumulative_profile = sys.getdxp() # 0s out our copy.
def merge_profile():
"""Reads sys.getdxp() and merges it into this module's cached copy.
We need this because sys.getdxp() 0s itself every time it's called."""
with _profile_lock:
new_profile = sys.getdxp()
if has_pairs(new_profile):
for first_inst in range(len(_cumulative_profile)):
for second_inst in range(len(_cumulative_profile[first_inst])):
_cumulative_profile[first_inst][second_inst] += (
new_profile[first_inst][second_inst])
else:
for inst in range(len(_cumulative_profile)):
_cumulative_profile[inst] += new_profile[inst]
def snapshot_profile():
"""Returns the cumulative execution profile until this call."""
with _profile_lock:
merge_profile()
return copy.deepcopy(_cumulative_profile)
def common_instructions(profile):
"""Returns the most common opcodes in order of descending frequency.
The result is a list of tuples of the form
(opcode, opname, # of occurrences)
"""
if has_pairs(profile) and profile:
inst_list = profile[-1]
else:
inst_list = profile
result = [(op, opcode.opname[op], count)
for op, count in enumerate(inst_list)
if count > 0]
result.sort(key=operator.itemgetter(2), reverse=True)
return result
def common_pairs(profile):
"""Returns the most common opcode pairs in order of descending frequency.
The result is a list of tuples of the form
((1st opcode, 2nd opcode),
(1st opname, 2nd opname),
# of occurrences of the pair)
"""
if not has_pairs(profile):
return []
result = [((op1, op2), (opcode.opname[op1], opcode.opname[op2]), count)
# Drop the row of single-op profiles with [:-1]
for op1, op1profile in enumerate(profile[:-1])
for op2, count in enumerate(op1profile)
if count > 0]
result.sort(key=operator.itemgetter(2), reverse=True)
return result
def render_common_pairs(profile=None):
"""Renders the most common opcode pairs to a string in order of
descending frequency.
The result is a series of lines of the form:
# of occurrences: ('1st opname', '2nd opname')
"""
if profile is None:
profile = snapshot_profile()
def seq():
for _, ops, count in common_pairs(profile):
yield "%s: %s\n" % (count, ops)
return ''.join(seq())