#!/usr/bin/env python
#
# Copyright 2008-2009 Jose Fonseca
#
# This program is free software: you can redistribute it and/or modify it
# under the terms of the GNU Lesser General Public License as published
# by the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with this program. If not, see .
#
"""Generate a dot graph from the output of several profilers."""
__author__ = "Jose Fonseca"
__version__ = "1.1"
import sys
import math
import os.path
import re
import textwrap
import optparse
import xml.parsers.expat
try:
# Debugging helper module
import debug
except ImportError:
pass
def times(x):
return "%u\xd7" % (x,)
def percentage(p):
return "%.02f%%" % (p*100.0,)
def add(a, b):
return a + b
def equal(a, b):
if a == b:
return a
else:
return None
def fail(a, b):
assert False
tol = 2 ** -23
def ratio(numerator, denominator):
try:
ratio = float(numerator)/float(denominator)
except ZeroDivisionError:
# 0/0 is undefined, but 1.0 yields more useful results
return 1.0
if ratio < 0.0:
if ratio < -tol:
sys.stderr.write('warning: negative ratio (%s/%s)\n' % (numerator, denominator))
return 0.0
if ratio > 1.0:
if ratio > 1.0 + tol:
sys.stderr.write('warning: ratio greater than one (%s/%s)\n' % (numerator, denominator))
return 1.0
return ratio
class UndefinedEvent(Exception):
"""Raised when attempting to get an event which is undefined."""
def __init__(self, event):
Exception.__init__(self)
self.event = event
def __str__(self):
return 'unspecified event %s' % self.event.name
class Event(object):
"""Describe a kind of event, and its basic operations."""
def __init__(self, name, null, aggregator, formatter = str):
self.name = name
self._null = null
self._aggregator = aggregator
self._formatter = formatter
def __eq__(self, other):
return self is other
def __hash__(self):
return id(self)
def null(self):
return self._null
def aggregate(self, val1, val2):
"""Aggregate two event values."""
assert val1 is not None
assert val2 is not None
return self._aggregator(val1, val2)
def format(self, val):
"""Format an event value."""
assert val is not None
return self._formatter(val)
CALLS = Event("Calls", 0, add, times)
SAMPLES = Event("Samples", 0, add)
SAMPLES2 = Event("Samples", 0, add)
TIME = Event("Time", 0.0, add, lambda x: '(' + str(x) + ')')
TIME_RATIO = Event("Time ratio", 0.0, add, lambda x: '(' + percentage(x) + ')')
TOTAL_TIME = Event("Total time", 0.0, fail)
TOTAL_TIME_RATIO = Event("Total time ratio", 0.0, fail, percentage)
class Object(object):
"""Base class for all objects in profile which can store events."""
def __init__(self, events=None):
if events is None:
self.events = {}
else:
self.events = events
def __hash__(self):
return id(self)
def __eq__(self, other):
return self is other
def __contains__(self, event):
return event in self.events
def __getitem__(self, event):
try:
return self.events[event]
except KeyError:
raise UndefinedEvent(event)
def __setitem__(self, event, value):
if value is None:
if event in self.events:
del self.events[event]
else:
self.events[event] = value
class Call(Object):
"""A call between functions.
There should be at most one call object for every pair of functions.
"""
def __init__(self, callee_id):
Object.__init__(self)
self.callee_id = callee_id
self.ratio = None
self.weight = None
class Function(Object):
"""A function."""
def __init__(self, id, name):
Object.__init__(self)
self.id = id
self.name = name
self.module = None
self.process = None
self.calls = {}
self.called = None
self.weight = None
self.cycle = None
def add_call(self, call):
if call.callee_id in self.calls:
sys.stderr.write('warning: overwriting call from function %s to %s\n' % (str(self.id), str(call.callee_id)))
self.calls[call.callee_id] = call
def get_call(self, callee_id):
if not callee_id in self.calls:
call = Call(callee_id)
call[SAMPLES] = 0
call[SAMPLES2] = 0
call[CALLS] = 0
self.calls[callee_id] = call
return self.calls[callee_id]
_parenthesis_re = re.compile(r'\([^()]*\)')
_angles_re = re.compile(r'<[^<>]*>')
_const_re = re.compile(r'\s+const$')
def stripped_name(self):
"""Remove extraneous information from C++ demangled function names."""
name = self.name
# Strip function parameters from name by recursively removing paired parenthesis
while True:
name, n = self._parenthesis_re.subn('', name)
if not n:
break
# Strip const qualifier
name = self._const_re.sub('', name)
# Strip template parameters from name by recursively removing paired angles
while True:
name, n = self._angles_re.subn('', name)
if not n:
break
return name
# TODO: write utility functions
def __repr__(self):
return self.name
class Cycle(Object):
"""A cycle made from recursive function calls."""
def __init__(self):
Object.__init__(self)
# XXX: Do cycles need an id?
self.functions = set()
def add_function(self, function):
assert function not in self.functions
self.functions.add(function)
# XXX: Aggregate events?
if function.cycle is not None:
for other in function.cycle.functions:
if function not in self.functions:
self.add_function(other)
function.cycle = self
class Profile(Object):
"""The whole profile."""
def __init__(self):
Object.__init__(self)
self.functions = {}
self.cycles = []
def add_function(self, function):
if function.id in self.functions:
sys.stderr.write('warning: overwriting function %s (id %s)\n' % (function.name, str(function.id)))
self.functions[function.id] = function
def add_cycle(self, cycle):
self.cycles.append(cycle)
def validate(self):
"""Validate the edges."""
for function in self.functions.values():
for callee_id in list(function.calls.keys()):
assert function.calls[callee_id].callee_id == callee_id
if callee_id not in self.functions:
sys.stderr.write('warning: call to undefined function %s from function %s\n' % (str(callee_id), function.name))
del function.calls[callee_id]
def find_cycles(self):
"""Find cycles using Tarjan's strongly connected components algorithm."""
# Apply the Tarjan's algorithm successively until all functions are visited
visited = set()
for function in self.functions.values():
if function not in visited:
self._tarjan(function, 0, [], {}, {}, visited)
cycles = []
for function in self.functions.values():
if function.cycle is not None and function.cycle not in cycles:
cycles.append(function.cycle)
self.cycles = cycles
if 0:
for cycle in cycles:
sys.stderr.write("Cycle:\n")
for member in cycle.functions:
sys.stderr.write("\tFunction %s\n" % member.name)
def _tarjan(self, function, order, stack, orders, lowlinks, visited):
"""Tarjan's strongly connected components algorithm.
See also:
- http://en.wikipedia.org/wiki/Tarjan's_strongly_connected_components_algorithm
"""
visited.add(function)
orders[function] = order
lowlinks[function] = order
order += 1
pos = len(stack)
stack.append(function)
for call in function.calls.values():
callee = self.functions[call.callee_id]
# TODO: use a set to optimize lookup
if callee not in orders:
order = self._tarjan(callee, order, stack, orders, lowlinks, visited)
lowlinks[function] = min(lowlinks[function], lowlinks[callee])
elif callee in stack:
lowlinks[function] = min(lowlinks[function], orders[callee])
if lowlinks[function] == orders[function]:
# Strongly connected component found
members = stack[pos:]
del stack[pos:]
if len(members) > 1:
cycle = Cycle()
for member in members:
cycle.add_function(member)
return order
def call_ratios(self, event):
# Aggregate for incoming calls
cycle_totals = {}
for cycle in self.cycles:
cycle_totals[cycle] = 0.0
function_totals = {}
for function in self.functions.values():
function_totals[function] = 0.0
for function in self.functions.values():
for call in function.calls.values():
if call.callee_id != function.id:
callee = self.functions[call.callee_id]
function_totals[callee] += call[event]
if callee.cycle is not None and callee.cycle is not function.cycle:
cycle_totals[callee.cycle] += call[event]
# Compute the ratios
for function in self.functions.values():
for call in function.calls.values():
assert call.ratio is None
if call.callee_id != function.id:
callee = self.functions[call.callee_id]
if callee.cycle is not None and callee.cycle is not function.cycle:
total = cycle_totals[callee.cycle]
else:
total = function_totals[callee]
call.ratio = ratio(call[event], total)
def integrate(self, outevent, inevent):
"""Propagate function time ratio allong the function calls.
Must be called after finding the cycles.
See also:
- http://citeseer.ist.psu.edu/graham82gprof.html
"""
# Sanity checking
assert outevent not in self
for function in self.functions.values():
assert outevent not in function
assert inevent in function
for call in function.calls.values():
assert outevent not in call
if call.callee_id != function.id:
assert call.ratio is not None
# Aggregate the input for each cycle
for cycle in self.cycles:
total = inevent.null()
for function in self.functions.values():
total = inevent.aggregate(total, function[inevent])
self[inevent] = total
# Integrate along the edges
total = inevent.null()
for function in self.functions.values():
total = inevent.aggregate(total, function[inevent])
self._integrate_function(function, outevent, inevent)
self[outevent] = total
def _integrate_function(self, function, outevent, inevent):
if function.cycle is not None:
return self._integrate_cycle(function.cycle, outevent, inevent)
else:
if outevent not in function:
total = function[inevent]
for call in function.calls.values():
if call.callee_id != function.id:
total += self._integrate_call(call, outevent, inevent)
function[outevent] = total
return function[outevent]
def _integrate_call(self, call, outevent, inevent):
assert outevent not in call
assert call.ratio is not None
callee = self.functions[call.callee_id]
subtotal = call.ratio *self._integrate_function(callee, outevent, inevent)
call[outevent] = subtotal
return subtotal
def _integrate_cycle(self, cycle, outevent, inevent):
if outevent not in cycle:
# Compute the outevent for the whole cycle
total = inevent.null()
for member in cycle.functions:
subtotal = member[inevent]
for call in member.calls.values():
callee = self.functions[call.callee_id]
if callee.cycle is not cycle:
subtotal += self._integrate_call(call, outevent, inevent)
total += subtotal
cycle[outevent] = total
# Compute the time propagated to callers of this cycle
callees = {}
for function in self.functions.values():
if function.cycle is not cycle:
for call in function.calls.values():
callee = self.functions[call.callee_id]
if callee.cycle is cycle:
try:
callees[callee] += call.ratio
except KeyError:
callees[callee] = call.ratio
for member in cycle.functions:
member[outevent] = outevent.null()
for callee, call_ratio in callees.items():
ranks = {}
call_ratios = {}
partials = {}
self._rank_cycle_function(cycle, callee, 0, ranks)
self._call_ratios_cycle(cycle, callee, ranks, call_ratios, set())
partial = self._integrate_cycle_function(cycle, callee, call_ratio, partials, ranks, call_ratios, outevent, inevent)
assert partial == max(partials.values())
assert not total or abs(1.0 - partial/(call_ratio*total)) <= 0.001
return cycle[outevent]
def _rank_cycle_function(self, cycle, function, rank, ranks):
if function not in ranks or ranks[function] > rank:
ranks[function] = rank
for call in function.calls.values():
if call.callee_id != function.id:
callee = self.functions[call.callee_id]
if callee.cycle is cycle:
self._rank_cycle_function(cycle, callee, rank + 1, ranks)
def _call_ratios_cycle(self, cycle, function, ranks, call_ratios, visited):
if function not in visited:
visited.add(function)
for call in function.calls.values():
if call.callee_id != function.id:
callee = self.functions[call.callee_id]
if callee.cycle is cycle:
if ranks[callee] > ranks[function]:
call_ratios[callee] = call_ratios.get(callee, 0.0) + call.ratio
self._call_ratios_cycle(cycle, callee, ranks, call_ratios, visited)
def _integrate_cycle_function(self, cycle, function, partial_ratio, partials, ranks, call_ratios, outevent, inevent):
if function not in partials:
partial = partial_ratio*function[inevent]
for call in function.calls.values():
if call.callee_id != function.id:
callee = self.functions[call.callee_id]
if callee.cycle is not cycle:
assert outevent in call
partial += partial_ratio*call[outevent]
else:
if ranks[callee] > ranks[function]:
callee_partial = self._integrate_cycle_function(cycle, callee, partial_ratio, partials, ranks, call_ratios, outevent, inevent)
call_ratio = ratio(call.ratio, call_ratios[callee])
call_partial = call_ratio*callee_partial
try:
call[outevent] += call_partial
except UndefinedEvent:
call[outevent] = call_partial
partial += call_partial
partials[function] = partial
try:
function[outevent] += partial
except UndefinedEvent:
function[outevent] = partial
return partials[function]
def aggregate(self, event):
"""Aggregate an event for the whole profile."""
total = event.null()
for function in self.functions.values():
try:
total = event.aggregate(total, function[event])
except UndefinedEvent:
return
self[event] = total
def ratio(self, outevent, inevent):
assert outevent not in self
assert inevent in self
for function in self.functions.values():
assert outevent not in function
assert inevent in function
function[outevent] = ratio(function[inevent], self[inevent])
for call in function.calls.values():
assert outevent not in call
if inevent in call:
call[outevent] = ratio(call[inevent], self[inevent])
self[outevent] = 1.0
def prune(self, node_thres, edge_thres):
"""Prune the profile"""
# compute the prune ratios
for function in self.functions.values():
try:
function.weight = function[TOTAL_TIME_RATIO]
except UndefinedEvent:
pass
for call in function.calls.values():
callee = self.functions[call.callee_id]
if TOTAL_TIME_RATIO in call:
# handle exact cases first
call.weight = call[TOTAL_TIME_RATIO]
else:
try:
# make a safe estimate
call.weight = min(function[TOTAL_TIME_RATIO], callee[TOTAL_TIME_RATIO])
except UndefinedEvent:
pass
# prune the nodes
for function_id in list(self.functions.keys()):
function = self.functions[function_id]
if function.weight is not None:
if function.weight < node_thres:
del self.functions[function_id]
# prune the egdes
for function in self.functions.values():
for callee_id in list(function.calls.keys()):
call = function.calls[callee_id]
if callee_id not in self.functions or call.weight is not None and call.weight < edge_thres:
del function.calls[callee_id]
def dump(self):
for function in self.functions.values():
sys.stderr.write('Function %s:\n' % (function.name,))
self._dump_events(function.events)
for call in function.calls.values():
callee = self.functions[call.callee_id]
sys.stderr.write(' Call %s:\n' % (callee.name,))
self._dump_events(call.events)
for cycle in self.cycles:
sys.stderr.write('Cycle:\n')
self._dump_events(cycle.events)
for function in cycle.functions:
sys.stderr.write(' Function %s\n' % (function.name,))
def _dump_events(self, events):
for event, value in events.items():
sys.stderr.write(' %s: %s\n' % (event.name, event.format(value)))
class Struct:
"""Masquerade a dictionary with a structure-like behavior."""
def __init__(self, attrs = None):
if attrs is None:
attrs = {}
self.__dict__['_attrs'] = attrs
def __getattr__(self, name):
try:
return self._attrs[name]
except KeyError:
raise AttributeError(name)
def __setattr__(self, name, value):
self._attrs[name] = value
def __str__(self):
return str(self._attrs)
def __repr__(self):
return repr(self._attrs)
class ParseError(Exception):
"""Raised when parsing to signal mismatches."""
def __init__(self, msg, line):
self.msg = msg
# TODO: store more source line information
self.line = line
def __str__(self):
return '%s: %r' % (self.msg, self.line)
class Parser:
"""Parser interface."""
def __init__(self):
pass
def parse(self):
raise NotImplementedError
class LineParser(Parser):
"""Base class for parsers that read line-based formats."""
def __init__(self, file):
Parser.__init__(self)
self._file = file
self.__line = None
self.__eof = False
self.line_no = 0
def readline(self):
line = self._file.readline()
if not line:
self.__line = ''
self.__eof = True
else:
self.line_no += 1
self.__line = line.rstrip('\r\n')
def lookahead(self):
assert self.__line is not None
return self.__line
def consume(self):
assert self.__line is not None
line = self.__line
self.readline()
return line
def eof(self):
assert self.__line is not None
return self.__eof
XML_ELEMENT_START, XML_ELEMENT_END, XML_CHARACTER_DATA, XML_EOF = list(range(4))
class XmlToken:
def __init__(self, type, name_or_data, attrs = None, line = None, column = None):
assert type in (XML_ELEMENT_START, XML_ELEMENT_END, XML_CHARACTER_DATA, XML_EOF)
self.type = type
self.name_or_data = name_or_data
self.attrs = attrs
self.line = line
self.column = column
def __str__(self):
if self.type == XML_ELEMENT_START:
return '<' + self.name_or_data + ' ...>'
if self.type == XML_ELEMENT_END:
return '' + self.name_or_data + '>'
if self.type == XML_CHARACTER_DATA:
return self.name_or_data
if self.type == XML_EOF:
return 'end of file'
assert 0
class XmlTokenizer:
"""Expat based XML tokenizer."""
def __init__(self, fp, skip_ws = True):
self.fp = fp
self.tokens = []
self.index = 0
self.final = False
self.skip_ws = skip_ws
self.character_pos = 0, 0
self.character_data = ''
self.parser = xml.parsers.expat.ParserCreate()
self.parser.StartElementHandler = self.handle_element_start
self.parser.EndElementHandler = self.handle_element_end
self.parser.CharacterDataHandler = self.handle_character_data
def handle_element_start(self, name, attributes):
self.finish_character_data()
line, column = self.pos()
token = XmlToken(XML_ELEMENT_START, name, attributes, line, column)
self.tokens.append(token)
def handle_element_end(self, name):
self.finish_character_data()
line, column = self.pos()
token = XmlToken(XML_ELEMENT_END, name, None, line, column)
self.tokens.append(token)
def handle_character_data(self, data):
if not self.character_data:
self.character_pos = self.pos()
self.character_data += data
def finish_character_data(self):
if self.character_data:
if not self.skip_ws or not self.character_data.isspace():
line, column = self.character_pos
token = XmlToken(XML_CHARACTER_DATA, self.character_data, None, line, column)
self.tokens.append(token)
self.character_data = ''
def __next__(self):
size = 16*1024
while self.index >= len(self.tokens) and not self.final:
self.tokens = []
self.index = 0
data = self.fp.read(size)
self.final = len(data) < size
try:
self.parser.Parse(data, self.final)
except xml.parsers.expat.ExpatError as e:
#if e.code == xml.parsers.expat.errors.XML_ERROR_NO_ELEMENTS:
if e.code == 3:
pass
else:
raise e
if self.index >= len(self.tokens):
line, column = self.pos()
token = XmlToken(XML_EOF, None, None, line, column)
else:
token = self.tokens[self.index]
self.index += 1
return token
def pos(self):
return self.parser.CurrentLineNumber, self.parser.CurrentColumnNumber
class XmlTokenMismatch(Exception):
def __init__(self, expected, found):
self.expected = expected
self.found = found
def __str__(self):
return '%u:%u: %s expected, %s found' % (self.found.line, self.found.column, str(self.expected), str(self.found))
class XmlParser(Parser):
"""Base XML document parser."""
def __init__(self, fp):
Parser.__init__(self)
self.tokenizer = XmlTokenizer(fp)
self.consume()
def consume(self):
self.token = next(self.tokenizer)
def match_element_start(self, name):
return self.token.type == XML_ELEMENT_START and self.token.name_or_data == name
def match_element_end(self, name):
return self.token.type == XML_ELEMENT_END and self.token.name_or_data == name
def element_start(self, name):
while self.token.type == XML_CHARACTER_DATA:
self.consume()
if self.token.type != XML_ELEMENT_START:
raise XmlTokenMismatch(XmlToken(XML_ELEMENT_START, name), self.token)
if self.token.name_or_data != name:
raise XmlTokenMismatch(XmlToken(XML_ELEMENT_START, name), self.token)
attrs = self.token.attrs
self.consume()
return attrs
def element_end(self, name):
while self.token.type == XML_CHARACTER_DATA:
self.consume()
if self.token.type != XML_ELEMENT_END:
raise XmlTokenMismatch(XmlToken(XML_ELEMENT_END, name), self.token)
if self.token.name_or_data != name:
raise XmlTokenMismatch(XmlToken(XML_ELEMENT_END, name), self.token)
self.consume()
def character_data(self, strip = True):
data = ''
while self.token.type == XML_CHARACTER_DATA:
data += self.token.name_or_data
self.consume()
if strip:
data = data.strip()
return data
class GprofParser(Parser):
"""Parser for GNU gprof output.
See also:
- Chapter "Interpreting gprof's Output" from the GNU gprof manual
http://sourceware.org/binutils/docs-2.18/gprof/Call-Graph.html#Call-Graph
- File "cg_print.c" from the GNU gprof source code
http://sourceware.org/cgi-bin/cvsweb.cgi/~checkout~/src/gprof/cg_print.c?rev=1.12&cvsroot=src
"""
def __init__(self, fp):
Parser.__init__(self)
self.fp = fp
self.functions = {}
self.cycles = {}
def readline(self):
line = self.fp.readline()
if not line:
sys.stderr.write('error: unexpected end of file\n')
sys.exit(1)
line = line.rstrip('\r\n')
return line
_int_re = re.compile(r'^\d+$')
_float_re = re.compile(r'^\d+\.\d+$')
def translate(self, mo):
"""Extract a structure from a match object, while translating the types in the process."""
attrs = {}
groupdict = mo.groupdict()
for name, value in groupdict.items():
if value is None:
value = None
elif self._int_re.match(value):
value = int(value)
elif self._float_re.match(value):
value = float(value)
attrs[name] = (value)
return Struct(attrs)
_cg_header_re = re.compile(
# original gprof header
r'^\s+called/total\s+parents\s*$|' +
r'^index\s+%time\s+self\s+descendents\s+called\+self\s+name\s+index\s*$|' +
r'^\s+called/total\s+children\s*$|' +
# GNU gprof header
r'^index\s+%\s+time\s+self\s+children\s+called\s+name\s*$'
)
_cg_ignore_re = re.compile(
# spontaneous
r'^\s+\s*$|'
# internal calls (such as "mcount")
r'^.*\((\d+)\)$'
)
_cg_primary_re = re.compile(
r'^\[(?P\d+)\]?' +
r'\s+(?P\d+\.\d+)' +
r'\s+(?P\d+\.\d+)' +
r'\s+(?P\d+\.\d+)' +
r'\s+(?:(?P\d+)(?:\+(?P\d+))?)?' +
r'\s+(?P\S.*?)' +
r'(?:\s+\d+)>)?' +
r'\s\[(\d+)\]$'
)
_cg_parent_re = re.compile(
r'^\s+(?P\d+\.\d+)?' +
r'\s+(?P\d+\.\d+)?' +
r'\s+(?P\d+)(?:/(?P\d+))?' +
r'\s+(?P\S.*?)' +
r'(?:\s+\d+)>)?' +
r'\s\[(?P\d+)\]$'
)
_cg_child_re = _cg_parent_re
_cg_cycle_header_re = re.compile(
r'^\[(?P\d+)\]?' +
r'\s+(?P\d+\.\d+)' +
r'\s+(?P\d+\.\d+)' +
r'\s+(?P\d+\.\d+)' +
r'\s+(?:(?P\d+)(?:\+(?P\d+))?)?' +
r'\s+\d+)\sas\sa\swhole>' +
r'\s\[(\d+)\]$'
)
_cg_cycle_member_re = re.compile(
r'^\s+(?P\d+\.\d+)?' +
r'\s+(?P\d+\.\d+)?' +
r'\s+(?P\d+)(?:\+(?P\d+))?' +
r'\s+(?P\S.*?)' +
r'(?:\s+\d+)>)?' +
r'\s\[(?P\d+)\]$'
)
_cg_sep_re = re.compile(r'^--+$')
def parse_function_entry(self, lines):
parents = []
children = []
while True:
if not lines:
sys.stderr.write('warning: unexpected end of entry\n')
line = lines.pop(0)
if line.startswith('['):
break
# read function parent line
mo = self._cg_parent_re.match(line)
if not mo:
if self._cg_ignore_re.match(line):
continue
sys.stderr.write('warning: unrecognized call graph entry: %r\n' % line)
else:
parent = self.translate(mo)
parents.append(parent)
# read primary line
mo = self._cg_primary_re.match(line)
if not mo:
sys.stderr.write('warning: unrecognized call graph entry: %r\n' % line)
return
else:
function = self.translate(mo)
while lines:
line = lines.pop(0)
# read function subroutine line
mo = self._cg_child_re.match(line)
if not mo:
if self._cg_ignore_re.match(line):
continue
sys.stderr.write('warning: unrecognized call graph entry: %r\n' % line)
else:
child = self.translate(mo)
children.append(child)
function.parents = parents
function.children = children
self.functions[function.index] = function
def parse_cycle_entry(self, lines):
# read cycle header line
line = lines[0]
mo = self._cg_cycle_header_re.match(line)
if not mo:
sys.stderr.write('warning: unrecognized call graph entry: %r\n' % line)
return
cycle = self.translate(mo)
# read cycle member lines
cycle.functions = []
for line in lines[1:]:
mo = self._cg_cycle_member_re.match(line)
if not mo:
sys.stderr.write('warning: unrecognized call graph entry: %r\n' % line)
continue
call = self.translate(mo)
cycle.functions.append(call)
self.cycles[cycle.cycle] = cycle
def parse_cg_entry(self, lines):
if lines[0].startswith("["):
self.parse_cycle_entry(lines)
else:
self.parse_function_entry(lines)
def parse_cg(self):
"""Parse the call graph."""
# skip call graph header
while not self._cg_header_re.match(self.readline()):
pass
line = self.readline()
while self._cg_header_re.match(line):
line = self.readline()
# process call graph entries
entry_lines = []
while line != '\014': # form feed
if line and not line.isspace():
if self._cg_sep_re.match(line):
self.parse_cg_entry(entry_lines)
entry_lines = []
else:
entry_lines.append(line)
line = self.readline()
def parse(self):
self.parse_cg()
self.fp.close()
profile = Profile()
profile[TIME] = 0.0
cycles = {}
for index in self.cycles.keys():
cycles[index] = Cycle()
for entry in self.functions.values():
# populate the function
function = Function(entry.index, entry.name)
function[TIME] = entry.self
if entry.called is not None:
function.called = entry.called
if entry.called_self is not None:
call = Call(entry.index)
call[CALLS] = entry.called_self
function.called += entry.called_self
# populate the function calls
for child in entry.children:
call = Call(child.index)
assert child.called is not None
call[CALLS] = child.called
if child.index not in self.functions:
# NOTE: functions that were never called but were discovered by gprof's
# static call graph analysis dont have a call graph entry so we need
# to add them here
missing = Function(child.index, child.name)
function[TIME] = 0.0
function.called = 0
profile.add_function(missing)
function.add_call(call)
profile.add_function(function)
if entry.cycle is not None:
try:
cycle = cycles[entry.cycle]
except KeyError:
sys.stderr.write('warning: entry missing\n' % entry.cycle)
cycle = Cycle()
cycles[entry.cycle] = cycle
cycle.add_function(function)
profile[TIME] = profile[TIME] + function[TIME]
for cycle in cycles.values():
profile.add_cycle(cycle)
# Compute derived events
profile.validate()
profile.ratio(TIME_RATIO, TIME)
profile.call_ratios(CALLS)
profile.integrate(TOTAL_TIME, TIME)
profile.ratio(TOTAL_TIME_RATIO, TOTAL_TIME)
return profile
class CallgrindParser(LineParser):
"""Parser for valgrind's callgrind tool.
See also:
- http://valgrind.org/docs/manual/cl-format.html
"""
_call_re = re.compile('^calls=\s*(\d+)\s+((\d+|\+\d+|-\d+|\*)\s+)+$')
def __init__(self, infile):
LineParser.__init__(self, infile)
# Textual positions
self.position_ids = {}
self.positions = {}
# Numeric positions
self.num_positions = 1
self.cost_positions = ['line']
self.last_positions = [0]
# Events
self.num_events = 0
self.cost_events = []
self.profile = Profile()
self.profile[SAMPLES] = 0
def parse(self):
# read lookahead
self.readline()
self.parse_key('version')
self.parse_key('creator')
while self.parse_part():
pass
if not self.eof():
sys.stderr.write('warning: line %u: unexpected line\n' % self.line_no)
sys.stderr.write('%s\n' % self.lookahead())
# compute derived data
self.profile.validate()
self.profile.find_cycles()
self.profile.ratio(TIME_RATIO, SAMPLES)
self.profile.call_ratios(CALLS)
self.profile.integrate(TOTAL_TIME_RATIO, TIME_RATIO)
return self.profile
def parse_part(self):
if not self.parse_header_line():
return False
while self.parse_header_line():
pass
if not self.parse_body_line():
return False
while self.parse_body_line():
pass
return True
def parse_header_line(self):
return \
self.parse_empty() or \
self.parse_comment() or \
self.parse_part_detail() or \
self.parse_description() or \
self.parse_event_specification() or \
self.parse_cost_line_def() or \
self.parse_cost_summary()
_detail_keys = set(('cmd', 'pid', 'thread', 'part'))
def parse_part_detail(self):
return self.parse_keys(self._detail_keys)
def parse_description(self):
return self.parse_key('desc') is not None
def parse_event_specification(self):
event = self.parse_key('event')
if event is None:
return False
return True
def parse_cost_line_def(self):
pair = self.parse_keys(('events', 'positions'))
if pair is None:
return False
key, value = pair
items = value.split()
if key == 'events':
self.num_events = len(items)
self.cost_events = items
if key == 'positions':
self.num_positions = len(items)
self.cost_positions = items
self.last_positions = [0]*self.num_positions
return True
def parse_cost_summary(self):
pair = self.parse_keys(('summary', 'totals'))
if pair is None:
return False
return True
def parse_body_line(self):
return \
self.parse_empty() or \
self.parse_comment() or \
self.parse_cost_line() or \
self.parse_position_spec() or \
self.parse_association_spec()
__subpos_re = r'(0x[0-9a-fA-F]+|\d+|\+\d+|-\d+|\*)'
_cost_re = re.compile(r'^' +
__subpos_re + r'( +' + __subpos_re + r')*' +
r'( +\d+)*' +
'$')
def parse_cost_line(self, calls=None):
line = self.lookahead().rstrip()
mo = self._cost_re.match(line)
if not mo:
return False
function = self.get_function()
if calls is None:
# Unlike other aspects, call object (cob) is relative not to the
# last call object, but to the caller's object (ob), so update it
# when processing a functions cost line
self.positions['cob'] = self.positions['ob']
values = line.split()
assert len(values) <= self.num_positions + self.num_events
positions = values[0 : self.num_positions]
events = values[self.num_positions : ]
events += ['0']*(self.num_events - len(events))
for i in range(self.num_positions):
position = positions[i]
if position == '*':
position = self.last_positions[i]
elif position[0] in '-+':
position = self.last_positions[i] + int(position)
elif position.startswith('0x'):
position = int(position, 16)
else:
position = int(position)
self.last_positions[i] = position
events = list(map(float, events))
if calls is None:
function[SAMPLES] += events[0]
self.profile[SAMPLES] += events[0]
else:
callee = self.get_callee()
callee.called += calls
try:
call = function.calls[callee.id]
except KeyError:
call = Call(callee.id)
call[CALLS] = calls
call[SAMPLES] = events[0]
function.add_call(call)
else:
call[CALLS] += calls
call[SAMPLES] += events[0]
self.consume()
return True
def parse_association_spec(self):
line = self.lookahead()
if not line.startswith('calls='):
return False
_, values = line.split('=', 1)
values = values.strip().split()
calls = int(values[0])
call_position = values[1:]
self.consume()
self.parse_cost_line(calls)
return True
_position_re = re.compile('^(?P[cj]?(?:ob|fl|fi|fe|fn))=\s*(?:\((?P\d+)\))?(?:\s*(?P.+))?')
_position_table_map = {
'ob': 'ob',
'fl': 'fl',
'fi': 'fl',
'fe': 'fl',
'fn': 'fn',
'cob': 'ob',
'cfl': 'fl',
'cfi': 'fl',
'cfe': 'fl',
'cfn': 'fn',
'jfi': 'fl',
}
_position_map = {
'ob': 'ob',
'fl': 'fl',
'fi': 'fl',
'fe': 'fl',
'fn': 'fn',
'cob': 'cob',
'cfl': 'cfl',
'cfi': 'cfl',
'cfe': 'cfl',
'cfn': 'cfn',
'jfi': 'jfi',
}
def parse_position_spec(self):
line = self.lookahead()
if line.startswith('jump=') or line.startswith('jcnd='):
self.consume()
return True
mo = self._position_re.match(line)
if not mo:
return False
position, id, name = mo.groups()
if id:
table = self._position_table_map[position]
if name:
self.position_ids[(table, id)] = name
else:
name = self.position_ids.get((table, id), '')
self.positions[self._position_map[position]] = name
self.consume()
return True
def parse_empty(self):
if self.eof():
return False
line = self.lookahead()
if line.strip():
return False
self.consume()
return True
def parse_comment(self):
line = self.lookahead()
if not line.startswith('#'):
return False
self.consume()
return True
_key_re = re.compile(r'^(\w+):')
def parse_key(self, key):
pair = self.parse_keys((key,))
if not pair:
return None
key, value = pair
return value
line = self.lookahead()
mo = self._key_re.match(line)
if not mo:
return None
key, value = line.split(':', 1)
if key not in keys:
return None
value = value.strip()
self.consume()
return key, value
def parse_keys(self, keys):
line = self.lookahead()
mo = self._key_re.match(line)
if not mo:
return None
key, value = line.split(':', 1)
if key not in keys:
return None
value = value.strip()
self.consume()
return key, value
def make_function(self, module, filename, name):
# FIXME: module and filename are not being tracked reliably
#id = '|'.join((module, filename, name))
id = name
try:
function = self.profile.functions[id]
except KeyError:
function = Function(id, name)
if module:
function.module = os.path.basename(module)
function[SAMPLES] = 0
function.called = 0
self.profile.add_function(function)
return function
def get_function(self):
module = self.positions.get('ob', '')
filename = self.positions.get('fl', '')
function = self.positions.get('fn', '')
return self.make_function(module, filename, function)
def get_callee(self):
module = self.positions.get('cob', '')
filename = self.positions.get('cfi', '')
function = self.positions.get('cfn', '')
return self.make_function(module, filename, function)
class PerfParser(LineParser):
"""Parser for linux perf callgraph output.
It expects output generated with
perf record -g
perf script | gprof2dot.py --format=perf
"""
def __init__(self, infile):
LineParser.__init__(self, infile)
self.profile = Profile()
def readline(self):
# Override LineParser.readline to ignore comment lines
while True:
LineParser.readline(self)
if self.eof() or not self.lookahead().startswith('#'):
break
def parse(self):
# read lookahead
self.readline()
profile = self.profile
profile[SAMPLES] = 0
while not self.eof():
self.parse_event()
# compute derived data
profile.validate()
profile.find_cycles()
profile.ratio(TIME_RATIO, SAMPLES)
profile.call_ratios(SAMPLES2)
profile.integrate(TOTAL_TIME_RATIO, TIME_RATIO)
return profile
def parse_event(self):
if self.eof():
return
line = self.consume()
assert line
callchain = self.parse_callchain()
if not callchain:
return
callee = callchain[0]
callee[SAMPLES] += 1
self.profile[SAMPLES] += 1
for caller in callchain[1:]:
try:
call = caller.calls[callee.id]
except KeyError:
call = Call(callee.id)
call[SAMPLES2] = 1
caller.add_call(call)
else:
call[SAMPLES2] += 1
callee = caller
def parse_callchain(self):
callchain = []
while self.lookahead():
function = self.parse_call()
if function is None:
break
callchain.append(function)
if self.lookahead() == '':
self.consume()
return callchain
call_re = re.compile(r'^\s+(?P[0-9a-fA-F]+)\s+(?P.*)\s+\((?P[^)]*)\)$')
def parse_call(self):
line = self.consume()
mo = self.call_re.match(line)
assert mo
if not mo:
return None
function_name = mo.group('symbol')
if not function_name:
function_name = mo.group('address')
module = mo.group('module')
function_id = function_name + ':' + module
try:
function = self.profile.functions[function_id]
except KeyError:
function = Function(function_id, function_name)
function.module = os.path.basename(module)
function[SAMPLES] = 0
self.profile.add_function(function)
return function
class OprofileParser(LineParser):
"""Parser for oprofile callgraph output.
See also:
- http://oprofile.sourceforge.net/doc/opreport.html#opreport-callgraph
"""
_fields_re = {
'samples': r'(\d+)',
'%': r'(\S+)',
'linenr info': r'(?P