# Class for profiling python code. rev 1.0 6/2/94
# Written by James Roskind
# Based on prior profile module by Sjoerd Mullender...
# which was hacked somewhat by: Guido van Rossum
"""Class for profiling Python code."""
# Copyright Disney Enterprises, Inc. All Rights Reserved.
# Licensed to PSF under a Contributor Agreement
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND,
# either express or implied. See the License for the specific language
# governing permissions and limitations under the License.
from optparse import OptionParser
__all__ = ["run", "runctx", "help", "Profile"]
# Sample timer for use with
#itimes = integer_timer # replace with C coded timer returning integers
#**************************************************************************
# The following are the static member functions for the profiler class
# Note that an instance of Profile() is *not* needed to call them.
#**************************************************************************
def run(statement, filename=None, sort=-1):
"""Run statement under profiler optionally saving results in filename
This function takes a single argument that can be passed to the
"exec" statement, and an optional file name. In all cases this
routine attempts to "exec" its first argument and gather profiling
statistics from the execution. If no file name is present, then this
function automatically prints a simple profiling report, sorted by the
standard name string (file/line/function-name) that is presented in
prof = prof.run(statement)
prof.dump_stats(filename)
return prof.print_stats(sort)
def runctx(statement, globals, locals, filename=None, sort=-1):
"""Run statement under profiler, supplying your own globals and locals,
optionally saving results in filename.
statement and filename have the same semantics as profile.run
prof = prof.runctx(statement, globals, locals)
prof.dump_stats(filename)
return prof.print_stats(sort)
# Backwards compatibility.
print "Documentation for the profile module can be found "
print "in the Python Library Reference, section 'The Python Profiler'."
def _get_time_times(timer=os.times):
# Using getrusage(3) is better than clock(3) if available:
# on some systems (e.g. FreeBSD), getrusage has a higher resolution
# Furthermore, on a POSIX system, returns microseconds, which
# wrap around after 36min.
resgetrusage = lambda: resource.getrusage(resource.RUSAGE_SELF)
def _get_time_resource(timer=resgetrusage):
self.cur is always a tuple. Each such tuple corresponds to a stack
frame that is currently active (self.cur[-2]). The following are the
definitions of its members. We use this external "parallel stack" to
avoid contaminating the program that we are profiling. (old profiler
used to write into the frames local dictionary!!) Derived classes
can change the definition of some entries, as long as they leave
[-2:] intact (frame and previous tuple). In case an internal error is
detected, the -3 element is used as the function name.
[ 0] = Time that needs to be charged to the parent frame's function.
It is used so that a function call will not have to access the
timing data for the parent frame.
[ 1] = Total time spent in this frame's function, excluding time in
subfunctions (this latter is tallied in cur[2]).
[ 2] = Total time spent in subfunctions, excluding time executing the
frame's function (this latter is tallied in cur[1]).
[-3] = Name of the function that corresponds to this frame.
[-2] = Actual frame that we correspond to (used to sync exception handling).
[-1] = Our parent 6-tuple (corresponds to frame.f_back).
Timing data for each function is stored as a 5-tuple in the dictionary
self.timings[]. The index is always the name stored in self.cur[-3].
The following are the definitions of the members:
[0] = The number of times this function was called, not counting direct
[1] = Number of times this function appears on the stack, minus one
[2] = Total time spent internal to this function
[3] = Cumulative time that this function was present on the stack. In
non-recursive functions, this is the total execution time from start
to finish of each invocation of a function, including time spent in
[4] = A dictionary indicating for each function name, the number of times
bias = 0 # calibration constant
def __init__(self, timer=None, bias=None):
self.bias = bias # Materialize in local dict for lookup speed.
self.timer = resgetrusage
self.dispatcher = self.trace_dispatch
self.get_time = _get_time_resource
elif hasattr(time, 'clock'):
self.timer = self.get_time = time.clock
self.dispatcher = self.trace_dispatch_i
elif hasattr(os, 'times'):
self.dispatcher = self.trace_dispatch
self.get_time = _get_time_times
self.timer = self.get_time = time.time
self.dispatcher = self.trace_dispatch_i
t = self.timer() # test out timer function
self.dispatcher = self.trace_dispatch_i
self.dispatcher = self.trace_dispatch
self.dispatcher = self.trace_dispatch_l
# This get_time() implementation needs to be defined
# here to capture the passed-in timer in the parameter
# list (for performance). Note that we can't assume
# the timer() result contains two values in all
def get_time_timer(timer=timer, sum=sum):
self.get_time = get_time_timer
self.simulate_call('profiler')
# Heavily optimized dispatch routine for os.times() timer
def trace_dispatch(self, frame, event, arg):
t = t[0] + t[1] - self.t - self.bias
self.c_func_name = arg.__name__
if self.dispatch[event](self, frame,t):
self.t = r[0] + r[1] - t # put back unrecorded delta
# Dispatch routine for best timer program (return = scalar, fastest if
# an integer but float works too -- and time.clock() relies on that).
def trace_dispatch_i(self, frame, event, arg):
t = timer() - self.t - self.bias
self.c_func_name = arg.__name__
if self.dispatch[event](self, frame, t):
self.t = timer() - t # put back unrecorded delta
# Dispatch routine for macintosh (timer returns time in ticks of
def trace_dispatch_mac(self, frame, event, arg):
t = timer()/60.0 - self.t - self.bias
self.c_func_name = arg.__name__
if self.dispatch[event](self, frame, t):
self.t = timer()/60.0 - t # put back unrecorded delta
# SLOW generic dispatch routine for timer returning lists of numbers
def trace_dispatch_l(self, frame, event, arg):
t = get_time() - self.t - self.bias
self.c_func_name = arg.__name__
if self.dispatch[event](self, frame, t):
self.t = get_time() - t # put back unrecorded delta
# In the event handlers, the first 3 elements of self.cur are unpacked
# into vrbls w/ 3-letter names. The last two characters are meant to be
# _pt self.cur[0] "parent time" time to be charged to parent frame
# _it self.cur[1] "internal time" time spent directly in the function
# _et self.cur[2] "external time" time spent in subfunctions
def trace_dispatch_exception(self, frame, t):
rpt, rit, ret, rfn, rframe, rcur = self.cur
if (rframe is not frame) and rcur:
return self.trace_dispatch_return(rframe, t)
self.cur = rpt, rit+t, ret, rfn, rframe, rcur
def trace_dispatch_call(self, frame, t):
if self.cur and frame.f_back is not self.cur[-2]:
rpt, rit, ret, rfn, rframe, rcur = self.cur
if not isinstance(rframe, Profile.fake_frame):
assert rframe.f_back is frame.f_back, ("Bad call", rfn,
self.trace_dispatch_return(rframe, 0)
assert (self.cur is None or \
frame.f_back is self.cur[-2]), ("Bad call",
fn = (fcode.co_filename, fcode.co_firstlineno, fcode.co_name)
self.cur = (t, 0, 0, fn, frame, self.cur)
cc, ns, tt, ct, callers = timings[fn]
timings[fn] = cc, ns + 1, tt, ct, callers
timings[fn] = 0, 0, 0, 0, {}
def trace_dispatch_c_call (self, frame, t):
fn = ("", 0, self.c_func_name)
self.cur = (t, 0, 0, fn, frame, self.cur)
cc, ns, tt, ct, callers = timings[fn]
timings[fn] = cc, ns+1, tt, ct, callers
timings[fn] = 0, 0, 0, 0, {}
def trace_dispatch_return(self, frame, t):
if frame is not self.cur[-2]:
assert frame is self.cur[-2].f_back, ("Bad return", self.cur[-3])
self.trace_dispatch_return(self.cur[-2], 0)
# Prefix "r" means part of the Returning or exiting frame.
# Prefix "p" means part of the Previous or Parent or older frame.
rpt, rit, ret, rfn, frame, rcur = self.cur
ppt, pit, pet, pfn, pframe, pcur = rcur
self.cur = ppt, pit + rpt, pet + frame_total, pfn, pframe, pcur
cc, ns, tt, ct, callers = timings[rfn]
# This is the only occurrence of the function on the stack.
# Else this is a (directly or indirectly) recursive call, and
# its cumulative time will get updated when the topmost call to
callers[pfn] = callers[pfn] + 1 # hack: gather more
# stats such as the amount of time added to ct courtesy
# of this specific call, and the contribution to cc
timings[rfn] = cc, ns - 1, tt + rit, ct, callers
"call": trace_dispatch_call,
"exception": trace_dispatch_exception,
"return": trace_dispatch_return,
"c_call": trace_dispatch_c_call,
"c_exception": trace_dispatch_return, # the C function returned
"c_return": trace_dispatch_return,
# The next few functions play with self.cmd. By carefully preloading
# our parallel stack, we can force the profiled result to include
# an arbitrary string as the name of the calling function.
# We use self.cmd as that string, and the resulting stats look
if self.cur[-1]: return # already set
def __init__(self, filename, line, name):
self.co_filename = filename
return repr((self.co_filename, self.co_line, self.co_name))
def __init__(self, code, prior):
def simulate_call(self, name):
code = self.fake_code('profile', 0, name)
frame = self.fake_frame(code, pframe)
self.dispatch['call'](self, frame, 0)
# collect stats from pending stack, including getting final
# timings for self.cmd frame.
def simulate_cmd_complete(self):
# We *can* cause assertion errors here if
# dispatch_trace_return checks for a frame match!
self.dispatch['return'](self, self.cur[-2], t)
def print_stats(self, sort=-1):
pstats.Stats(self).strip_dirs().sort_stats(sort). \
def dump_stats(self, file):
marshal.dump(self.stats, f)
self.simulate_cmd_complete()
def snapshot_stats(self):
for func, (cc, ns, tt, ct, callers) in self.timings.iteritems():
for callcnt in callers.itervalues():
self.stats[func] = cc, nc, tt, ct, callers
# The following two methods can be called by clients to use
# a profiler to profile a statement, given as a string.
return self.runctx(cmd, dict, dict)
def runctx(self, cmd, globals, locals):
sys.setprofile(self.dispatcher)
exec cmd in globals, locals
# This method is more useful to profile a single function call.
def runcall(self, func, *args, **kw):
sys.setprofile(self.dispatcher)
#******************************************************************
# The following calculates the overhead for using a profiler. The
# problem is that it takes a fair amount of time for the profiler
# to stop the stopwatch (from the time it receives an event).
# Similarly, there is a delay from the time that the profiler
# re-starts the stopwatch before the user's code really gets to
# continue. The following code tries to measure the difference on
# Note that this difference is only significant if there are a lot of
# events, and relatively little user code per event. For example,
# code with small functions will typically benefit from having the
# profiler calibrated for the current platform. This *could* be
# done on the fly during init() time, but it is not worth the
# effort. Also note that if too large a value specified, then
# execution time on some functions will actually appear as a
# negative number. It is *normal* for some functions (with very
# low call counts) to have such negative stats, even if the
# calibration figure is "correct."
# One alternative to profile-time calibration adjustments (i.e.,
# adding in the magic little delta during each event) is to track
# more carefully the number of events (and cumulatively, the number
# of events during sub functions) that are seen. If this were
# done, then the arithmetic could be done after the fact (i.e., at
# display time). Currently, we track only call/return events.
# These values can be deduced by examining the callees and callers
# vectors for each functions. Hence we *can* almost correct the
# internal time figure at print time (note that we currently don't
# track exception event processing counts). Unfortunately, there
# is currently no similar information for cumulative sub-function
# time. It would not be hard to "get all this info" at profiler
# time. Specifically, we would have to extend the tuples to keep
# counts of this in each frame, and then extend the defs of timing
# tuples to include the significant two figures. I'm a bit fearful
# that this additional feature will slow the heavily optimized
# event/time ratio (i.e., the profiler would run slower, fur a very
# low "value added" feature.)
#**************************************************************
def calibrate(self, m, verbose=0):
if self.__class__ is not Profile:
It is recommended that you Edit text format, this type of Fix handles quite a lot in one request
