""" Profiling hooks This module contains a couple of decorators (`profile` and `coverage`) that can be used to wrap functions and/or methods to produce profiles and line coverage reports. There's a third convenient decorator (`timecall`) that measures the duration of function execution without the extra profiling overhead. Usage example (Python 2.4 or newer):: from profilehooks import profile, coverage @profile # or @coverage def fn(n): if n < 2: return 1 else: return n * fn(n-1) print fn(42) Usage example (Python 2.3 or older):: from profilehooks import profile, coverage def fn(n): if n < 2: return 1 else: return n * fn(n-1) # Now wrap that function in a decorator fn = profile(fn) # or coverage(fn) print fn(42) Reports for all thusly decorated functions will be printed to sys.stdout on program termination. You can alternatively request for immediate reports for each call by passing immediate=True to the profile decorator. There's also a @timecall decorator for printing the time to sys.stderr every time a function is called, when you just want to get a rough measure instead of a detailed (but costly) profile. Caveats A thread on python-dev convinced me that hotshot produces bogus numbers. See http://mail.python.org/pipermail/python-dev/2005-November/058264.html I don't know what will happen if a decorated function will try to call another decorated function. All decorators probably need to explicitly support nested profiling (currently TraceFuncCoverage is the only one that supports this, while HotShotFuncProfile has support for recursive functions.) Profiling with hotshot creates temporary files (*.prof for profiling, *.cprof for coverage) in the current directory. These files are not cleaned up. Exception: when you specify a filename to the profile decorator (to store the pstats.Stats object for later inspection), the temporary file will be the filename you specified with '.raw' appended at the end. Coverage analysis with hotshot seems to miss some executions resulting in lower line counts and some lines errorneously marked as never executed. For this reason coverage analysis now uses trace.py which is slower, but more accurate. Copyright (c) 2004--2008 Marius Gedminas Copyright (c) 2007 Hanno Schlichting Copyright (c) 2008 Florian Schulze Released under the MIT licence since December 2006: Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. (Previously it was distributed under the GNU General Public Licence.) """ # $Id: profilehooks.py 29 2010-08-13 16:29:20Z mg $ __author__ = "Marius Gedminas (marius@gedmin.as)" __copyright__ = "Copyright 2004-2009 Marius Gedminas" __license__ = "MIT" __version__ = "1.4" __date__ = "2009-03-31" import atexit import inspect import sys import re # For profiling from profile import Profile import pstats # For hotshot profiling (inaccurate!) try: import hotshot import hotshot.stats except ImportError: hotshot = None # For trace.py coverage import trace # For hotshot coverage (inaccurate!; uses undocumented APIs; might break) if hotshot is not None: import _hotshot import hotshot.log # For cProfile profiling (best) try: import cProfile except ImportError: cProfile = None # For timecall import time # registry of available profilers AVAILABLE_PROFILERS = {} def profile(fn=None, skip=0, filename=None, immediate=False, dirs=False, sort=None, entries=40, profiler=('cProfile', 'profile', 'hotshot')): """Mark `fn` for profiling. If `skip` is > 0, first `skip` calls to `fn` will not be profiled. If `immediate` is False, profiling results will be printed to sys.stdout on program termination. Otherwise results will be printed after each call. If `dirs` is False only the name of the file will be printed. Otherwise the full path is used. `sort` can be a list of sort keys (defaulting to ['cumulative', 'time', 'calls']). The following ones are recognized:: 'calls' -- call count 'cumulative' -- cumulative time 'file' -- file name 'line' -- line number 'module' -- file name 'name' -- function name 'nfl' -- name/file/line 'pcalls' -- call count 'stdname' -- standard name 'time' -- internal time `entries` limits the output to the first N entries. `profiler` can be used to select the preferred profiler, or specify a sequence of them, in order of preference. The default is ('cProfile'. 'profile', 'hotshot'). If `filename` is specified, the profile stats will be stored in the named file. You can load them pstats.Stats(filename). Usage:: def fn(...): ... fn = profile(fn, skip=1) If you are using Python 2.4, you should be able to use the decorator syntax:: @profile(skip=3) def fn(...): ... or just :: @profile def fn(...): ... """ if fn is None: # @profile() syntax -- we are a decorator maker def decorator(fn): return profile(fn, skip=skip, filename=filename, immediate=immediate, dirs=dirs, sort=sort, entries=entries, profiler=profiler) return decorator # @profile syntax -- we are a decorator. if isinstance(profiler, str): profiler = [profiler] for p in profiler: if p in AVAILABLE_PROFILERS: profiler_class = AVAILABLE_PROFILERS[p] break else: raise ValueError('only these profilers are available: %s' % ', '.join(AVAILABLE_PROFILERS)) fp = profiler_class(fn, skip=skip, filename=filename, immediate=immediate, dirs=dirs, sort=sort, entries=entries) # fp = HotShotFuncProfile(fn, skip=skip, filename=filename, ...) # or HotShotFuncProfile # We cannot return fp or fp.__call__ directly as that would break method # definitions, instead we need to return a plain function. def new_fn(*args, **kw): return fp(*args, **kw) new_fn.__doc__ = fn.__doc__ new_fn.__name__ = fn.__name__ new_fn.__dict__ = fn.__dict__ new_fn.__module__ = fn.__module__ return new_fn def coverage(fn): """Mark `fn` for line coverage analysis. Results will be printed to sys.stdout on program termination. Usage:: def fn(...): ... fn = coverage(fn) If you are using Python 2.4, you should be able to use the decorator syntax:: @coverage def fn(...): ... """ fp = TraceFuncCoverage(fn) # or HotShotFuncCoverage # We cannot return fp or fp.__call__ directly as that would break method # definitions, instead we need to return a plain function. def new_fn(*args, **kw): return fp(*args, **kw) new_fn.__doc__ = fn.__doc__ new_fn.__name__ = fn.__name__ new_fn.__dict__ = fn.__dict__ new_fn.__module__ = fn.__module__ return new_fn def coverage_with_hotshot(fn): """Mark `fn` for line coverage analysis. Uses the 'hotshot' module for fast coverage analysis. BUG: Produces inaccurate results. See the docstring of `coverage` for usage examples. """ fp = HotShotFuncCoverage(fn) # We cannot return fp or fp.__call__ directly as that would break method # definitions, instead we need to return a plain function. def new_fn(*args, **kw): return fp(*args, **kw) new_fn.__doc__ = fn.__doc__ new_fn.__name__ = fn.__name__ new_fn.__dict__ = fn.__dict__ new_fn.__module__ = fn.__module__ return new_fn class FuncProfile(object): """Profiler for a function (uses profile).""" # This flag is shared between all instances in_profiler = False Profile = Profile def __init__(self, fn, skip=0, filename=None, immediate=False, dirs=False, sort=None, entries=40): """Creates a profiler for a function. Every profiler has its own log file (the name of which is derived from the function name). FuncProfile registers an atexit handler that prints profiling information to sys.stderr when the program terminates. """ self.fn = fn self.skip = skip self.filename = filename self.immediate = immediate self.dirs = dirs self.sort = sort or ('cumulative', 'time', 'calls') if isinstance(self.sort, str): self.sort = (self.sort, ) self.entries = entries self.reset_stats() atexit.register(self.atexit) def __call__(self, *args, **kw): """Profile a singe call to the function.""" self.ncalls += 1 if self.skip > 0: self.skip -= 1 self.skipped += 1 return self.fn(*args, **kw) if FuncProfile.in_profiler: # handle recursive calls return self.fn(*args, **kw) # You cannot reuse the same profiler for many calls and accumulate # stats that way. :-/ profiler = self.Profile() try: FuncProfile.in_profiler = True return profiler.runcall(self.fn, *args, **kw) finally: FuncProfile.in_profiler = False self.stats.add(profiler) if self.immediate: self.print_stats() self.reset_stats() def print_stats(self): """Print profile information to sys.stdout.""" funcname = self.fn.__name__ filename = self.fn.func_code.co_filename lineno = self.fn.func_code.co_firstlineno print print "*** PROFILER RESULTS ***" print "%s (%s:%s)" % (funcname, filename, lineno) print "function called %d times" % self.ncalls, if self.skipped: print "(%d calls not profiled)" % self.skipped else: print print stats = self.stats if self.filename: stats.dump_stats(self.filename) if not self.dirs: stats.strip_dirs() stats.sort_stats(*self.sort) stats.print_stats(self.entries) def reset_stats(self): """Reset accumulated profiler statistics.""" # Note: not using self.Profile, since pstats.Stats() fails then self.stats = pstats.Stats(Profile()) self.ncalls = 0 self.skipped = 0 def atexit(self): """Stop profiling and print profile information to sys.stdout. This function is registered as an atexit hook. """ if not self.immediate: self.print_stats() AVAILABLE_PROFILERS['profile'] = FuncProfile if cProfile is not None: class CProfileFuncProfile(FuncProfile): """Profiler for a function (uses cProfile).""" Profile = cProfile.Profile AVAILABLE_PROFILERS['cProfile'] = CProfileFuncProfile if hotshot is not None: class HotShotFuncProfile(object): """Profiler for a function (uses hotshot).""" # This flag is shared between all instances in_profiler = False def __init__(self, fn, skip=0, filename=None): """Creates a profiler for a function. Every profiler has its own log file (the name of which is derived from the function name). HotShotFuncProfile registers an atexit handler that prints profiling information to sys.stderr when the program terminates. The log file is not removed and remains there to clutter the current working directory. """ self.fn = fn self.filename = filename if self.filename: self.logfilename = filename + ".raw" else: self.logfilename = fn.__name__ + ".prof" self.profiler = hotshot.Profile(self.logfilename) self.ncalls = 0 self.skip = skip self.skipped = 0 atexit.register(self.atexit) def __call__(self, *args, **kw): """Profile a singe call to the function.""" self.ncalls += 1 if self.skip > 0: self.skip -= 1 self.skipped += 1 return self.fn(*args, **kw) if HotShotFuncProfile.in_profiler: # handle recursive calls return self.fn(*args, **kw) try: HotShotFuncProfile.in_profiler = True return self.profiler.runcall(self.fn, *args, **kw) finally: HotShotFuncProfile.in_profiler = False def atexit(self): """Stop profiling and print profile information to sys.stderr. This function is registered as an atexit hook. """ self.profiler.close() funcname = self.fn.__name__ filename = self.fn.func_code.co_filename lineno = self.fn.func_code.co_firstlineno print print "*** PROFILER RESULTS ***" print "%s (%s:%s)" % (funcname, filename, lineno) print "function called %d times" % self.ncalls, if self.skipped: print "(%d calls not profiled)" % self.skipped else: print print stats = hotshot.stats.load(self.logfilename) # hotshot.stats.load takes ages, and the .prof file eats megabytes, but # a saved stats object is small and fast if self.filename: stats.dump_stats(self.filename) # it is best to save before strip_dirs stats.strip_dirs() stats.sort_stats('cumulative', 'time', 'calls') stats.print_stats(40) AVAILABLE_PROFILERS['hotshot'] = HotShotFuncProfile class HotShotFuncCoverage: """Coverage analysis for a function (uses _hotshot). HotShot coverage is reportedly faster than trace.py, but it appears to have problems with exceptions; also line counts in coverage reports are generally lower from line counts produced by TraceFuncCoverage. Is this my bug, or is it a problem with _hotshot? """ def __init__(self, fn): """Creates a profiler for a function. Every profiler has its own log file (the name of which is derived from the function name). HotShotFuncCoverage registers an atexit handler that prints profiling information to sys.stderr when the program terminates. The log file is not removed and remains there to clutter the current working directory. """ self.fn = fn self.logfilename = fn.__name__ + ".cprof" self.profiler = _hotshot.coverage(self.logfilename) self.ncalls = 0 atexit.register(self.atexit) def __call__(self, *args, **kw): """Profile a singe call to the function.""" self.ncalls += 1 return self.profiler.runcall(self.fn, args, kw) def atexit(self): """Stop profiling and print profile information to sys.stderr. This function is registered as an atexit hook. """ self.profiler.close() funcname = self.fn.__name__ filename = self.fn.func_code.co_filename lineno = self.fn.func_code.co_firstlineno print print "*** COVERAGE RESULTS ***" print "%s (%s:%s)" % (funcname, filename, lineno) print "function called %d times" % self.ncalls print fs = FuncSource(self.fn) reader = hotshot.log.LogReader(self.logfilename) for what, (filename, lineno, funcname), tdelta in reader: if filename != fs.filename: continue if what == hotshot.log.LINE: fs.mark(lineno) if what == hotshot.log.ENTER: # hotshot gives us the line number of the function definition # and never gives us a LINE event for the first statement in # a function, so if we didn't perform this mapping, the first # statement would be marked as never executed if lineno == fs.firstlineno: lineno = fs.firstcodelineno fs.mark(lineno) reader.close() print fs class TraceFuncCoverage: """Coverage analysis for a function (uses trace module). HotShot coverage analysis is reportedly faster, but it appears to have problems with exceptions. """ # Shared between all instances so that nested calls work tracer = trace.Trace(count=True, trace=False, ignoredirs=[sys.prefix, sys.exec_prefix]) # This flag is also shared between all instances tracing = False def __init__(self, fn): """Creates a profiler for a function. Every profiler has its own log file (the name of which is derived from the function name). TraceFuncCoverage registers an atexit handler that prints profiling information to sys.stderr when the program terminates. The log file is not removed and remains there to clutter the current working directory. """ self.fn = fn self.logfilename = fn.__name__ + ".cprof" self.ncalls = 0 atexit.register(self.atexit) def __call__(self, *args, **kw): """Profile a singe call to the function.""" self.ncalls += 1 if TraceFuncCoverage.tracing: return self.fn(*args, **kw) try: TraceFuncCoverage.tracing = True return self.tracer.runfunc(self.fn, *args, **kw) finally: TraceFuncCoverage.tracing = False def atexit(self): """Stop profiling and print profile information to sys.stderr. This function is registered as an atexit hook. """ funcname = self.fn.__name__ filename = self.fn.func_code.co_filename lineno = self.fn.func_code.co_firstlineno print print "*** COVERAGE RESULTS ***" print "%s (%s:%s)" % (funcname, filename, lineno) print "function called %d times" % self.ncalls print fs = FuncSource(self.fn) for (filename, lineno), count in self.tracer.counts.items(): if filename != fs.filename: continue fs.mark(lineno, count) print fs never_executed = fs.count_never_executed() if never_executed: print "%d lines were not executed." % never_executed class FuncSource: """Source code annotator for a function.""" blank_rx = re.compile(r"^\s*finally:\s*(#.*)?$") def __init__(self, fn): self.fn = fn self.filename = inspect.getsourcefile(fn) self.source, self.firstlineno = inspect.getsourcelines(fn) self.sourcelines = {} self.firstcodelineno = self.firstlineno self.find_source_lines() def find_source_lines(self): """Mark all executable source lines in fn as executed 0 times.""" strs = trace.find_strings(self.filename) lines = trace.find_lines_from_code(self.fn.func_code, strs) self.firstcodelineno = sys.maxint for lineno in lines: self.firstcodelineno = min(self.firstcodelineno, lineno) self.sourcelines.setdefault(lineno, 0) if self.firstcodelineno == sys.maxint: self.firstcodelineno = self.firstlineno def mark(self, lineno, count=1): """Mark a given source line as executed count times. Multiple calls to mark for the same lineno add up. """ self.sourcelines[lineno] = self.sourcelines.get(lineno, 0) + count def count_never_executed(self): """Count statements that were never executed.""" lineno = self.firstlineno counter = 0 for line in self.source: if self.sourcelines.get(lineno) == 0: if not self.blank_rx.match(line): counter += 1 lineno += 1 return counter def __str__(self): """Return annotated source code for the function.""" lines = [] lineno = self.firstlineno for line in self.source: counter = self.sourcelines.get(lineno) if counter is None: prefix = ' ' * 7 elif counter == 0: if self.blank_rx.match(line): prefix = ' ' * 7 else: prefix = '>' * 6 + ' ' else: prefix = '%5d: ' % counter lines.append(prefix + line) lineno += 1 return ''.join(lines) def timecall(fn=None, immediate=True, timer=time.time): """Wrap `fn` and print its execution time. Example:: @timecall def somefunc(x, y): time.sleep(x * y) somefunc(2, 3) will print the time taken by somefunc on every call. If you want just a summary at program termination, use @timecall(immediate=False) You can also choose a timing method other than the default ``time.time()``, e.g.: @timecall(timer=time.clock) """ if fn is None: # @timecall() syntax -- we are a decorator maker def decorator(fn): return timecall(fn, immediate=immediate, timer=timer) return decorator # @timecall syntax -- we are a decorator. fp = FuncTimer(fn, immediate=immediate, timer=timer) # We cannot return fp or fp.__call__ directly as that would break method # definitions, instead we need to return a plain function. def new_fn(*args, **kw): return fp(*args, **kw) new_fn.__doc__ = fn.__doc__ new_fn.__name__ = fn.__name__ new_fn.__dict__ = fn.__dict__ new_fn.__module__ = fn.__module__ return new_fn class FuncTimer(object): def __init__(self, fn, immediate, timer): self.fn = fn self.ncalls = 0 self.totaltime = 0 self.immediate = immediate self.timer = timer if not immediate: atexit.register(self.atexit) def __call__(self, *args, **kw): """Profile a singe call to the function.""" fn = self.fn timer = self.timer self.ncalls += 1 try: start = timer() return fn(*args, **kw) finally: duration = timer() - start self.totaltime += duration if self.immediate: funcname = fn.__name__ filename = fn.func_code.co_filename lineno = fn.func_code.co_firstlineno print >> sys.stderr, "\n %s (%s:%s):\n %.3f seconds\n" % ( funcname, filename, lineno, duration) def atexit(self): if not self.ncalls: return funcname = self.fn.__name__ filename = self.fn.func_code.co_filename lineno = self.fn.func_code.co_firstlineno print ("\n %s (%s:%s):\n" " %d calls, %.3f seconds (%.3f seconds per call)\n" % ( funcname, filename, lineno, self.ncalls, self.totaltime, self.totaltime / self.ncalls))