perf.py

#!/usr/bin/env python

"""
Stolen from Unladen Swallow
(http://unladen-swallow.googlecode.com/svn/tests/perf.py). I'm just using
some parts of this -- stats results, t-tests, etc. -- but I've snagged
the whole file because I'm lazy. --JKM

----

Tool for comparing the performance of two Python implementations.

Typical usage looks like

./perf.py -b 2to3,django control/python experiment/python

This will run the 2to3 and Django template benchmarks, using `control/python`
as the baseline and `experiment/python` as the experiment. The --fast and
--rigorous options can be used to vary the duration/accuracy of the run. Run
--help to get a full list of options that can be passed to -b.

perf.py will run Student's two-tailed T test on the benchmark results at the 95%
confidence level to indicate whether the observed difference is statistically
significant.

Omitting the -b option will result in the default group of benchmarks being run
This currently consists of: 2to3, django, nbody, rietveld, slowspitfire,
slowpickle, slowunpickle, spambayes. Omitting -b is the same as specifying
`-b default`.

To run every benchmark perf.py knows about, use `-b all`. To see a full list of
all available benchmarks, use `--help`.

Negative benchmarks specifications are also supported: `-b -2to3` will run every
benchmark in the default group except for 2to3 (this is the same as
`-b default,-2to3`). `-b all,-django` will run all benchmarks except the Django
templates benchmark. Negative groups (e.g., `-b -default`) are not supported.
Positive benchmarks are parsed before the negative benchmarks are subtracted.

If --track_memory is passed, perf.py will continuously sample the benchmark's
memory usage, then give you the maximum usage and a link to a Google Chart of
the benchmark's memory usage over time. This currently only works on Linux
2.6.16 and higher or Windows with PyWin32. Because --track_memory introduces
performance jitter while collecting memory measurements, only memory usage is
reported in the final report.

If --args is passed, it specifies extra arguments to pass to the test
python binaries. For example,
  perf.py --args="-A -B,-C -D" base_python changed_python
will run benchmarks like
  base_python -A -B the_benchmark.py
  changed_python -C -D the_benchmark.py
while
  perf.py --args="-A -B" base_python changed_python
will pass the same arguments to both pythons:
  base_python -A -B the_benchmark.py
  changed_python -A -B the_benchmark.py
"""

from __future__ import division, with_statement

__author__ = "jyasskin@google.com (Jeffrey Yasskin)"

import contextlib
import logging
import math
import optparse
import os
import os.path
import platform
import re
import shutil
import subprocess
import sys
import tempfile
import time
import threading
import urllib2
try:
    import multiprocessing
except ImportError:
    multiprocessing = None
try:
    import win32api
    import win32con
    import win32process
    import pywintypes
except ImportError:
    win32api = None


info = logging.info


def avg(seq):
    return sum(seq) / float(len(seq))


def SampleStdDev(seq):
    """Compute the standard deviation of a sample.

    Args:
        seq: the numeric input data sequence.

    Returns:
        The standard deviation as a float.
    """
    mean = avg(seq)
    squares = ((x - mean) ** 2 for x in seq)
    return math.sqrt(sum(squares) / (len(seq) - 1))


# A table of 95% confidence intervals for a two-tailed t distribution, as a
# function of the degrees of freedom. For larger degrees of freedom, we
# approximate. While this may look less elegant than simply calculating the
# critical value, those calculations suck. Look at
# http://www.math.unb.ca/~knight/utility/t-table.htm if you need more values.
T_DIST_95_CONF_LEVELS = [0, 12.706, 4.303, 3.182, 2.776,
                         2.571, 2.447, 2.365, 2.306, 2.262,
                         2.228, 2.201, 2.179, 2.160, 2.145,
                         2.131, 2.120, 2.110, 2.101, 2.093,
                         2.086, 2.080, 2.074, 2.069, 2.064,
                         2.060, 2.056, 2.052, 2.048, 2.045,
                         2.042]


def TDist95ConfLevel(df):
    """Approximate the 95% confidence interval for Student's T distribution.

    Given the degrees of freedom, returns an approximation to the 95%
    confidence interval for the Student's T distribution.

    Args:
        df: An integer, the number of degrees of freedom.

    Returns:
        A float.
    """
    df = int(round(df))
    highest_table_df = len(T_DIST_95_CONF_LEVELS)
    if df >= 200: return 1.960
    if df >= 100: return 1.984
    if df >= 80: return 1.990
    if df >= 60: return 2.000
    if df >= 50: return 2.009
    if df >= 40: return 2.021
    if df >= highest_table_df:
        return T_DIST_95_CONF_LEVELS[highest_table_df - 1]
    return T_DIST_95_CONF_LEVELS[df]


def PooledSampleVariance(sample1, sample2):
    """Find the pooled sample variance for two samples.

    Args:
        sample1: one sample.
        sample2: the other sample.

    Returns:
        Pooled sample variance, as a float.
    """
    deg_freedom = len(sample1) + len(sample2) - 2
    mean1 = avg(sample1)
    squares1 = ((x - mean1) ** 2 for x in sample1)
    mean2 = avg(sample2)
    squares2 = ((x - mean2) ** 2 for x in sample2)

    return (sum(squares1) + sum(squares2)) / float(deg_freedom)


def TScore(sample1, sample2):
    """Calculate a t-test score for the difference between two samples.

    Args:
        sample1: one sample.
        sample2: the other sample.

    Returns:
        The t-test score, as a float.
    """
    assert len(sample1) == len(sample2)
    error = PooledSampleVariance(sample1, sample2) / len(sample1)
    return (avg(sample1) - avg(sample2)) / math.sqrt(error * 2)


def IsSignificant(sample1, sample2):
    """Determine whether two samples differ significantly.

    This uses a Student's two-sample, two-tailed t-test with alpha=0.95.

    Args:
        sample1: one sample.
        sample2: the other sample.

    Returns:
        (significant, t_score) where significant is a bool indicating whether
        the two samples differ significantly; t_score is the score from the
        two-sample T test.
    """
    deg_freedom = len(sample1) + len(sample2) - 2
    critical_value = TDist95ConfLevel(deg_freedom)
    t_score = TScore(sample1, sample2)
    return (abs(t_score) >= critical_value, t_score)


### Code to parse Linux /proc/%d/smaps files.
### See http://bmaurer.blogspot.com/2006/03/memory-usage-with-smaps.html for
### a quick introduction to smaps.

def _ParseSmapsData(smaps_data):
    """Parse the contents of a Linux 2.6 smaps file.

    Args:
        smaps_data: the smaps file contents, as a string.

    Returns:
        The size of the process's private data, in kilobytes.
    """
    total = 0
    for line in smaps_data.splitlines():
        # Include both Private_Clean and Private_Dirty sections.
        if line.startswith("Private_"):
            parts = line.split()
            total += int(parts[1])
    return total


def _ReadSmapsFile(pid):
    """Read the Linux smaps file for a pid.

    Args:
        pid: the process id to retrieve smaps data for.

    Returns:
        The data from the smaps file, as a string.

    Raises:
        IOError if the smaps file for the given pid could not be found.
    """
    with open("/proc/%d/smaps" % pid) as f:
        return f.read()


# Code to sample memory usage on Win32

def _GetWin32MemorySample(process_handle):
    """Gets the amount of memory in use by a process on Win32

    Args:
        process_handle: handle to the process to get the memory usage for

    Returns:
        The size of the process's private data, in kilobytes
    """
    pmi = win32process.GetProcessMemoryInfo(process_handle)
    return pmi["PagefileUsage"] // 1024


@contextlib.contextmanager
def _OpenWin32Process(pid):
    """Open a process on Win32 and close it when done

    Args:
        pid: the process id of the process to open

    Yields:
        A handle to the process

    Raises:
        pywintypes.error if the process does not exist or the user
            does not have sufficient privileges to open it

    Example:
        with _OpenWin32Process(pid) as process_handle:
            ...
    """
    h = win32api.OpenProcess(
            win32con.PROCESS_QUERY_INFORMATION | win32con.PROCESS_VM_READ,
            0,
            pid)
    try:
        yield h
    finally:
        win32api.CloseHandle(h)


def CanGetMemoryUsage():
    """Returns True if MemoryUsageFuture is supported on this platform."""
    if win32api:
        try:
            with _OpenWin32Process(win32process.GetCurrentProcessId()):
                return True
        except pywintypes.error:
            pass

    try:
        _ReadSmapsFile(pid=1)
    except IOError:
        pass
    else:
        return True

    return False


class MemoryUsageFuture(threading.Thread):
    """Continuously sample a process's memory usage for its lifetime.

    Example:
        future = MemoryUsageFuture(some_pid)
        ...
        usage = future.GetMemoryUsage()
        print max(usage)

    Note that calls to GetMemoryUsage() will block until the process exits.
    """

    def __init__(self, pid):
        super(MemoryUsageFuture, self).__init__()
        self._pid = pid
        self._usage = []
        self._done = threading.Event()
        self.start()

    def run(self):
        if win32api:
            with _OpenWin32Process(self._pid) as process_handle:
                while (win32process.GetExitCodeProcess(process_handle) ==
                       win32con.STILL_ACTIVE):
                    sample = _GetWin32MemorySample(process_handle)
                    self._usage.append(sample)
                    time.sleep(0.001)
        else:
            while True:
                try:
                    sample = _ParseSmapsData(_ReadSmapsFile(self._pid))
                    self._usage.append(sample)
                except IOError:
                    # Once the process exits, its smaps file will go away,
                    # leading _ReadSmapsFile() to raise IOError.
                    break
        self._done.set()

    def GetMemoryUsage(self):
        """Get the memory usage over time for the process being sampled.

        This will block until the process has exited.

        Returns:
            A list of all memory usage samples, in kilobytes.
        """
        self._done.wait()
        return self._usage


class RawData(object):
    """Raw data from a benchmark run.

    Attributes:
        runtimes: list of floats, one per iteration.
        mem_usage: list of ints, memory usage in kilobytes.
        inst_output: output from Unladen's --with-instrumentation build. This is
            the empty string if there was no instrumentation output.
    """

    def __init__(self, runtimes, mem_usage, inst_output=""):
        self.runtimes = runtimes
        self.mem_usage = mem_usage
        self.inst_output = inst_output


class BenchmarkResult(object):
    """An object representing data from a succesful benchmark run."""

    def __init__(self, min_base, min_changed, delta_min, avg_base,
                 avg_changed, delta_avg, t_msg, std_base, std_changed,
                 delta_std, timeline_link):
        self.min_base      = min_base
        self.min_changed   = min_changed
        self.delta_min     = delta_min
        self.avg_base      = avg_base
        self.avg_changed   = avg_changed
        self.delta_avg     = delta_avg
        self.t_msg         = t_msg
        self.std_base      = std_base
        self.std_changed   = std_changed
        self.delta_std     = delta_std
        self.timeline_link = timeline_link

    def get_timeline(self):
        if self.timeline_link is None:
            return ""
        return "\nTimeline: %(timeline_link)s" % self.__dict__

    def __str__(self):
        return (("Min: %(min_base)f -> %(min_changed)f:" +
                 " %(delta_min)s\n" +
                 "Avg: %(avg_base)f -> %(avg_changed)f:" +
                 " %(delta_avg)s\n" + self.t_msg +
                 "Stddev: %(std_base).5f -> %(std_changed).5f:" +
                 " %(delta_std)s" + self.get_timeline())
                 % self.__dict__)


class BenchmarkError(object):
    """Object representing the error from a failed benchmark run."""

    def __init__(self, e):
        self.msg = str(e)

    def __str__(self):
        return self.msg


class MemoryUsageResult(object):
    """Memory usage data from a successful benchmark run."""

    def __init__(self, max_base, max_changed, delta_max, timeline_link):
        self.max_base      = max_base
        self.max_changed   = max_changed
        self.delta_max     = delta_max
        self.timeline_link = timeline_link

    def get_usage_over_time(self):
        if self.timeline_link is None:
            return ""
        return "\nUsage over time: %(timeline_link)s" % self.__dict__

    def __str__(self):
        return (("Mem max: %(max_base).3f -> %(max_changed).3f:" +
                 " %(delta_max)s" + self.get_usage_over_time())
                 % self.__dict__)


class SimpleBenchmarkResult(object):
    """Object representing result data from a successful benchmark run."""

    def __init__(self, base_time, changed_time, time_delta):
        self.base_time    = base_time
        self.changed_time = changed_time
        self.time_delta   = time_delta

    def __str__(self):
        return ("%(base_time)f -> %(changed_time)f: %(time_delta)s"
                % self.__dict__)


class InstrumentationResult(object):
    """Object respresenting a --diff_instrumentation result."""

    def __init__(self, inst_diff, options):
        self.inst_diff = inst_diff
        self._control_label = options.control_label
        self._experiment_label = options.experiment_label

    def __str__(self):
        if not self.inst_diff:
            return "No difference in instrumentation"
        output = []
        for header, (control, exp) in self.inst_diff.items():
            output.append(header)
            output.append(self._control_label)
            output.append(control or "No data")
            output.append("")
            output.append(self._experiment_label)
            output.append(exp or "No data")
            output.append("\n")
        return "\n".join(output).strip()


def CompareMemoryUsage(base_usage, changed_usage, options):
    """Like CompareMultipleRuns, but for memory usage.

    Args:
        base_usage: list of the memory usage numbers for the control.
        changed_usage: list of the memory usage numbers for the experiment.
        options: optparse.Values instance.

    Returns:
        A MemoryUsageResult object.
    """
    max_base, max_changed = max(base_usage), max(changed_usage)
    delta_max = QuantityDelta(max_base, max_changed)

    if options.disable_timelines:
        chart_link = None
    else:
        chart_link = GetChart(SummarizeData(base_usage),
                              SummarizeData(changed_usage),
                              options,
                              title=options.benchmark_name,
                              y_label="Memory+(kb)")

    return MemoryUsageResult(max_base, max_changed, delta_max, chart_link)


### Utility functions


def _FormatPerfDataForTable(base_label, changed_label, results):
    """Prepare performance data for tabular output.

    Args:
        base_label: label for the control binary.
        changed_label: label for the experimental binary.
        results: iterable of (bench_name, result) 2-tuples where bench_name is
            the name of the benchmark being reported; and result is a
            BenchmarkResult object.

    Returns:
        A list of 6-tuples, where each tuple corresponds to a row in the output
        table, and each item in the tuples corresponds to a cell in the output
        table.
    """
    table = [("Benchmark", base_label, changed_label, "Change", "Significance",
              "Timeline")]

    for (bench_name, result) in results:
        table.append((bench_name,
                      # Limit the precision for conciseness in the table.
                      str(round(result.avg_base, 2)),
                      str(round(result.avg_changed, 2)),
                      result.delta_avg,
                      result.t_msg.strip(),
                      result.timeline_link))

    return table


def _FormatMemoryUsageForTable(base_label, changed_label, results):
    """Prepare memory usage data for tabular output.

    Args:
        base_label: label for the control binary.
        changed_label: label for the experimental binary.
        results: iterable of (bench_name, result) 2-tuples where bench_name is
            the name of the benchmark being reported; and result is a
            MemoryUsageResult object.

    Returns:
        A list of 5-tuples, where each tuple corresponds to a row in the output
        table, and each item in the tuples corresponds to a cell in the output
        table.
    """
    table = [("Benchmark", base_label, changed_label, "Change", "Timeline")]

    for (bench_name, result) in results:
        table.append((bench_name,
                      # We don't care about fractional kilobytes.
                      str(int(result.max_base)),
                      str(int(result.max_changed)),
                      result.delta_max,
                      result.timeline_link))

    return table


def FormatOutputAsTable(base_label, changed_label, results):
    """Format a benchmark result in a PEP-fiendly ASCII-art table.

    Args:
        base_label: label to use for the baseline binary.
        changed_label: label to use for the experimental binary.
        results: list of (bench_name, result) 2-tuples, where bench_name is the
            name of the just-run benchmark; and result is a BenchmarkResult
            object.

    Returns:
        A string holding the desired ASCII-art table.
    """
    if isinstance(results[0][1], BenchmarkResult):
        table = _FormatPerfDataForTable(base_label, changed_label, results)
    elif isinstance(results[0][1], MemoryUsageResult):
        table = _FormatMemoryUsageForTable(base_label, changed_label, results)
    else:
        raise TypeError("Unknown result type: %r" % type(results[0][1]))

    col_widths = [0] * len(table[0])
    for row in table:
        for col, val in enumerate(row):
            col_widths[col] = max(col_widths[col], len(val))

    outside_line = "+"
    header_sep_line = "+"
    for width in col_widths:
        width += 2  # Compensate for the left and right padding spaces.
        outside_line += "-" * width + "+"
        header_sep_line += "=" * width + "+"

    output = [outside_line]
    for row_i, row in enumerate(table):
        output_row = []
        for col_i, val in enumerate(row):
            output_row.append("| " + val.ljust(col_widths[col_i]) + " ")
        output.append("".join(output_row) + "|")
        if row_i > 0:
            output.append(outside_line)

    output.insert(2, "".join(header_sep_line))
    return "\n".join(output)


def _SegmentInstrumentation(inst_output):
    """Cut --with-instrumentation output into its component sections.

    Instrumentation sections are separated by two newlines, and begin with a
    header that ends in a colon and a newline.

    Args:
        inst_output: text holding full --with-instrumentation output.

    Returns:
        Dict mapping string section headers to section output text.
    """
    if not inst_output:
        return {}

    sections = {}
    text_sections = [s.strip() for s in inst_output.split("\n\n")]
    for section in text_sections:
        header, lines = section.split("\n", 1)
        if header.endswith(":"):
            sections[header] = lines
    return sections


def DiffInstrumentation(control_inst_output, exp_inst_output):
    """Compare the instrumentation output from two Unladen Swallow binaries.

    These binaries should have been configured with Unladen's
    --with-instrumentation flag.

    Args:
        control_inst_output: string; the control binary's instrumentation data.
        exp_inst_output: string; the experimental binary's instrumentation data.

    Returns:
        Dict mapping section headers to (control, exp) 2-tuples, where `control`
        is the output section from control binary, and `exp` is the output
        section from the experimental binary. If either `control` or `exp` is
        the empty string, that binary did not emit the section.
    """
    control_sections = _SegmentInstrumentation(control_inst_output)
    exp_sections = _SegmentInstrumentation(exp_inst_output)
    control_keys = set(control_sections)
    exp_keys = set(exp_sections)

    diff = {}
    for section in (control_keys - exp_keys):
        diff[section] = (control_sections[section], "")
    for section in (exp_keys - control_keys):
        diff[section] = ("", exp_sections[section])
    for section in (exp_keys & control_keys):
        if control_sections[section] != exp_sections[section]:
            diff[section] = (control_sections[section], exp_sections[section])
    return diff


def SimpleBenchmark(benchmark_function, base_python, changed_python, options,
                    *args, **kwargs):
    """Abstract out the body for most simple benchmarks.

    Example usage:
        def BenchmarkSomething(*args, **kwargs):
            return SimpleBenchmark(MeasureSomething, *args, **kwargs)

    The *args, **kwargs style is recommended so as to minimize the number of
    places that have to be changed if we update benchmark arguments.

    Args:
        benchmark_function: callback that takes (python_path, options) and
            returns a RawData instance.
        base_python: path to the reference Python binary.
        changed_python: path to the experimental Python binary.
        options: optparse.Values instance.
        *args, **kwargs: will be passed through to benchmark_function.

    Returns:
        A BenchmarkResult object if the benchmark runs succeeded.
        A BenchmarkError object if either benchmark run failed.
    """
    try:
        changed_data = benchmark_function(changed_python, options,
                                          *args, **kwargs)
        base_data = benchmark_function(base_python, options,
                                       *args, **kwargs)
    except subprocess.CalledProcessError, e:
        return BenchmarkError(e)

    return CompareBenchmarkData(base_data, changed_data, options)


def _FormatData(num):
    return str(round(num, 2))

def GetChart(base_data, changed_data, options, title, y_label,
             chart_margin=100):
    """Build a Google Chart API URL for the given data.

    Args:
        base_data: data points for the base binary.
        changed_data: data points for the changed binary.
        options: optparse.Values instance.
        title: title for the chart.
        y_label: label for Y axis on the chart.
        chart_margin: optional integer margin to add/sub from the max/min.

    Returns:
        Google Chart API URL as a string; or None, if options.disable_timelines
        is true.
    """
    if options.disable_timelines:
        return None
    # We use these to scale the graph.
    max_data = max(max(base_data), max(changed_data)) + chart_margin
    min_data = min(min(base_data), min(changed_data)) - chart_margin
    if min_data < 0:
        min_data = 0
    # Google-bound data, formatted as desired by the Chart API.
    data_for_google = (",".join(map(_FormatData, base_data)) + "|" +
                       ",".join(map(_FormatData, changed_data)))

    # Come up with labels for the X axis; not too many, though, or they'll be
    # unreadable.
    max_len = max(len(base_data), len(changed_data))
    if max_len <= 20:
        points = range(1, max_len + 1)
    else:
        points = SummarizeData(range(1, max_len + 1), points=5)
        if points[0] != 1:
            points.insert(0, 1)
    x_axis_labels = "".join("|%d" % i for i in points)

    # Parameters for the Google Chart API. See
    # http://code.google.com/apis/chart/ for more details.
    # cht=lc: line graph with visible axes.
    # chs: dimensions of the graph, in pixels.
    # chdl: labels for the graph lines.
    # chco: colors for the graph lines.
    # chds: minimum and maximum values for the vertical axis.
    # chxr: minimum and maximum values for the vertical axis labels.
    # chd=t: the data sets, |-separated.
    # chxt: which axes to draw.
    # chxl: labels for the axes.
    # chtt: chart title, using + for space and | for line breaks
    control_label = options.control_label
    experiment_label = options.experiment_label
    title = title.replace(' ', '+').replace('\n', '|')
    raw_url = ("http://chart.apis.google.com/chart?cht=lc&chs=700x400&"
               "chxt=x,y,x,y&"
               "chxr=1,%(min_data)s,%(max_data)s&chco=FF0000,0000FF&"
               "chdl=%(control_label)s|%(experiment_label)s&"
               "chds=%(min_data)s,%(max_data)s&chd=t:%(data_for_google)s&"
               "chxl=0:%(x_axis_labels)s|2:||Iteration|3:||%(y_label)s&"
               "chtt=%(title)s"
               % locals())
    return ShortenUrl(raw_url)


def ShortenUrl(url):
    """Shorten a given URL using tinyurl.com.

    Args:
        url: url to shorten.

    Returns:
        Shorter url. If tinyurl.com is not available, returns the original
        url unaltered.
    """
    tinyurl_api = "http://tinyurl.com/api-create.php?url="
    try:
        url = urllib2.urlopen(tinyurl_api + url).read()
    except urllib2.URLError:
        info("failed to call out to tinyurl.com")
    return url


def SummarizeData(data, points=100, summary_func=max):
    """Summarize a large data set using a smaller number of points.

    This will divide up the original data set into `points` windows,
    using `summary_func` to summarize each window into a single point.

    Args:
        data: the original data set, as a list.
        points: optional; how many summary points to take. Default is 100.
        summary_func: optional; function to use when summarizing each window.
            Default is the max() built-in.

    Returns:
        List of summary data points.
    """
    window_size = int(math.ceil(len(data) / points))
    if window_size == 1:
        return data

    summary_points = []
    start = 0
    while start < len(data):
        end = min(start + window_size, len(data))
        summary_points.append(summary_func(data[start:end]))
        start = end
    return summary_points


@contextlib.contextmanager
def ChangeDir(new_cwd):
    former_cwd = os.getcwd()
    os.chdir(new_cwd)
    try:
        yield
    finally:
        os.chdir(former_cwd)


def RemovePycs():
    if sys.platform == "win32":
        for root, dirs, files in os.walk('.'):
            for name in files:
                if name.endswith('.pyc') or name.endswith('.pyo'):
                    os.remove(os.path.join(root, name))
    else:
        subprocess.check_call(["find", ".", "-name", "*.py[co]",
                               "-exec", "rm", "-f", "{}", ";"])


def Relative(path):
    return os.path.join(os.path.dirname(__file__), path)


def LogCall(command):
    command = map(str, command)
    info("Running %s", " ".join(command))
    return command


try:
    import resource
except ImportError:
    # Approximate child time using wall clock time.
    def GetChildUserTime():
        return time.time()
else:
    def GetChildUserTime():
        return resource.getrusage(resource.RUSAGE_CHILDREN).ru_utime


@contextlib.contextmanager
def TemporaryFilename(prefix):
    fd, name = tempfile.mkstemp(prefix=prefix)
    os.close(fd)
    try:
        yield name
    finally:
        os.remove(name)


def TimeDelta(old, new):
    if old == 0 or new == 0:
        return "incomparable (one result was zero)"
    if new > old:
        return "%.4fx slower" % (new / old)
    elif new < old:
        return "%.4fx faster" % (old / new)
    else:
        return "no change"


def QuantityDelta(old, new):
    if old == 0 or new == 0:
        return "incomparable (one result was zero)"
    if new > old:
        return "%.4fx larger" % (new / old)
    elif new < old:
        return "%.4fx smaller" % (old / new)
    else:
        return "no change"


def BuildEnv(env=None, inherit_env=[]):
    """Massage an environment variables dict for the host platform.

    Massaging performed (in this order):
    - Add any variables named in inherit_env.
    - Copy PYTHONPATH to JYTHONPATH to support Jython.
    - Win32 requires certain env vars to be set.

    Args:
        env: optional; environment variables dict. If this is omitted, start
            with an empty environment.
        inherit_env: optional; iterable of strings, each the name of an
            environment variable to inherit from os.environ.

    Returns:
        A copy of `env`, possibly with modifications.
    """
    if env is None:
        env = {}
    fixed_env = env.copy()
    for varname in inherit_env:
        fixed_env[varname] = os.environ[varname]
    if "PYTHONPATH" in fixed_env:
        fixed_env["JYTHONPATH"] = fixed_env["PYTHONPATH"]
    if sys.platform == "win32":
        # Win32 requires certain environment variables be present
        for k in ("COMSPEC", "SystemRoot"):
            if k in os.environ and k not in fixed_env:
                fixed_env[k] = os.environ[k]
    return fixed_env


def CompareMultipleRuns(base_times, changed_times, options):
    """Compare multiple control vs experiment runs of the same benchmark.

    Args:
        base_times: iterable of float times (control).
        changed_times: iterable of float times (experiment).
        options: optparse.Values instance.

    Returns:
        A BenchmarkResult object, summarizing the difference between the two
        runs; or a SimpleBenchmarkResult object, if there was only one data
        point per run.
    """
    assert len(base_times) == len(changed_times)
    if len(base_times) == 1:
        # With only one data point, we can't do any of the interesting stats
        # below.
        base_time, changed_time = base_times[0], changed_times[0]
        time_delta = TimeDelta(base_time, changed_time)
        return SimpleBenchmarkResult(base_time, changed_time, time_delta)

    # Create a chart showing iteration times over time. We round the times so
    # as not to exceed the GET limit for Google's chart server.
    timeline_link = GetChart(SummarizeData(base_times),
                             SummarizeData(changed_times),
                             options,
                             title=options.benchmark_name,
                             y_label="Time+(secs)",
                             chart_margin=1)

    base_times = sorted(base_times)
    changed_times = sorted(changed_times)

    min_base, min_changed = base_times[0], changed_times[0]
    avg_base, avg_changed = avg(base_times), avg(changed_times)
    std_base = SampleStdDev(base_times)
    std_changed = SampleStdDev(changed_times)
    delta_min = TimeDelta(min_base, min_changed)
    delta_avg = TimeDelta(avg_base, avg_changed)
    delta_std = QuantityDelta(std_base, std_changed)

    t_msg = "Not significant\n"
    significant, t_score = IsSignificant(base_times, changed_times)
    if significant:
        t_msg = "Significant (t=%f)\n" % t_score

    return BenchmarkResult(min_base, min_changed, delta_min, avg_base,
                           avg_changed, delta_avg, t_msg, std_base,
                           std_changed, delta_std, timeline_link)


def CompareBenchmarkData(base_data, exp_data, options):
    """Compare performance and memory usage.

    Args:
        base_data: RawData instance for the control binary.
        exp_data: RawData instance for the experimental binary.
        options: optparse.Values instance.

    Returns:
        Something that implements a __str__() method:

        - BenchmarkResult: summarizes the difference between the two runs.
        - SimpleBenchmarkResult: if there was only one data point per run.
        - InstrumentationResult: if --diff_instrumentation was given.
        - MemoryUsageResult: if --track_memory was given.
        - BenchmarkError: if something went wrong.
    """
    # We suppress performance data when running with --track_memory or
    # --diff_instrumentation.
    if options.track_memory:
        if base_data.mem_usage is not None:
            assert exp_data.mem_usage is not None
            return CompareMemoryUsage(base_data.mem_usage, exp_data.mem_usage,
                                      options)
        return BencharkError("Benchmark does not report memory usage yet")
    if options.diff_instrumentation:
        inst_diff = DiffInstrumentation(base_data.inst_output,
                                        exp_data.inst_output)
        return InstrumentationResult(inst_diff, options)

    return CompareMultipleRuns(base_data.runtimes, exp_data.runtimes, options)


def CallAndCaptureOutput(command, env=None, track_memory=False, inherit_env=[]):
    """Run the given command, capturing stdout.

    Args:
        command: the command to run as a list, one argument per element.
        env: optional; environment variables to set.
        track_memory: optional; whether to continuously sample the subprocess's
            memory usage.
        inherit_env: optional; iterable of strings, each the name of an
            environment variable to inherit from os.environ.

    Returns:
        (stdout, stderr, mem_usage), where stdout is the captured stdout as a
        string; stderr is the captured stderr as a string; mem_usage is a list
        of memory usage samples in kilobytes (if track_memory is False,
        mem_usage is None).

    Raises:
        RuntimeError: if the command failed. The value of the exception will
        be the error message from the command.
    """
    mem_usage = None
    subproc = subprocess.Popen(LogCall(command),
                               stdout=subprocess.PIPE,
                               stderr=subprocess.PIPE,
                               env=BuildEnv(env, inherit_env))
    if track_memory:
        future = MemoryUsageFuture(subproc.pid)
    stdout, stderr = subproc.communicate()
    if subproc.returncode != 0:
        raise RuntimeError("Benchmark died: " + stderr)
    if track_memory:
        mem_usage = future.GetMemoryUsage()
    return stdout, stderr, mem_usage


def MeasureGeneric(python, options, bm_path, bm_env=None,
                   extra_args=[], iteration_scaling=1):
    """Abstract measurement function for Unladen's bm_* scripts.

    Based on the values of options.fast/rigorous, will pass -n {5,50,100} to
    the benchmark script. MeasureGeneric takes care of parsing out the running
    times from the memory usage data.

    Args:
        python: start of the argv list for running Python.
        options: optparse.Values instance.
        bm_path: path to the benchmark script.
        bm_env: optional environment dict. If this is unspecified or None,
            use an empty enviroment.
        extra_args: optional list of command line args to be given to the
            benchmark script.
        iteration_scaling: optional multiple by which to scale the -n argument
            to the benchmark.

    Returns:
        RawData instance.
    """
    if bm_env is None:
        bm_env = {}

    trials = 50
    if options.rigorous:
        trials = 100
    elif options.fast:
        trials = 5
    trials = max(1, int(trials * iteration_scaling))

    RemovePycs()
    command = python + [bm_path, "-n", trials] + extra_args
    output = CallAndCaptureOutput(command, bm_env,
                                  track_memory=options.track_memory,
                                  inherit_env=options.inherit_env)
    stdout, stderr, mem_usage = output
    times = [float(line) for line in stdout.splitlines()]
    return RawData(times, mem_usage, inst_output=stderr)


### Benchmarks

class PyBenchBenchmarkResult(object):

    def __init__(self, min_base, min_changed, delta_min,
                 avg_base, avg_changed, delta_avg):
        self.min_base = min_base
        self.min_changed = min_changed
        self.delta_min = delta_min
        self.avg_base = avg_base
        self.avg_changed = avg_changed
        self.delta_avg = delta_avg

    def __str__(self):
        return (("Min: %(min_base)d -> %(min_changed)d: %(delta_min)s\n" +
                 "Avg: %(avg_base)d -> %(avg_changed)d: %(delta_avg)s")
                % self.__dict__)


_PY_BENCH_TOTALS_LINE = re.compile("""
    Totals:\s+(?P<min_base>\d+)ms\s+
    (?P<min_changed>\d+)ms\s+
    \S+\s+  # Percent change, which we re-compute
    (?P<avg_base>\d+)ms\s+
    (?P<avg_changed>\d+)ms\s+
    \S+  # Second percent change, also re-computed
    """, re.X)

def MungePyBenchTotals(line):
    m = _PY_BENCH_TOTALS_LINE.search(line)
    if m:
        min_base, min_changed, avg_base, avg_changed = map(float, m.group(
            "min_base", "min_changed", "avg_base", "avg_changed"))
        delta_min = TimeDelta(min_base, min_changed)
        delta_avg = TimeDelta(avg_base, avg_changed)
        return PyBenchBenchmarkResult(min_base, min_changed, delta_min,
                                      avg_base, avg_changed, delta_avg)
    return BenchmarkError(line)


def BM_PyBench(base_python, changed_python, options):
    if options.track_memory:
        return BenchmarkError("Benchmark does not report memory usage yet")

    warp = "10"
    if options.rigorous:
        warp = "1"
    if options.fast:
        warp = "100"

    PYBENCH_PATH = Relative("performance/pybench/pybench.py")
    PYBENCH_ENV = BuildEnv(inherit_env=options.inherit_env)

    try:
        with contextlib.nested(open(os.devnull, "wb"),
                               TemporaryFilename(prefix="baseline."),
                               TemporaryFilename(prefix="changed.")
                               ) as (dev_null, base_pybench, changed_pybench):
            RemovePycs()
            subprocess.check_call(LogCall(changed_python + [
                                           PYBENCH_PATH,
                                           "-w", warp,
                                           "-f", changed_pybench,
                                           ]), stdout=dev_null,
                                           env=PYBENCH_ENV)
            RemovePycs()
            subprocess.check_call(LogCall(base_python + [
                                           PYBENCH_PATH,
                                           "-w", warp,
                                           "-f", base_pybench,
                                           ]), stdout=dev_null,
                                           env=PYBENCH_ENV)
            comparer = subprocess.Popen(base_python + [
                                         PYBENCH_PATH,
                                         "--debug",
                                         "-s", base_pybench,
                                         "-c", changed_pybench,
                                         ], stdout=subprocess.PIPE,
                                         stderr=subprocess.PIPE,
                                         env=PYBENCH_ENV)
            result, err = comparer.communicate()
            if comparer.returncode != 0:
                return BenchmarkError("pybench died: " + err)
    except subprocess.CalledProcessError, e:
        return BenchmarkError(e)

    if options.verbose:
        return BenchmarkError(result)
    else:
        for line in result.splitlines():
            if line.startswith("Totals:"):
                return MungePyBenchTotals(line)
        # The format's wrong...
        return BenchmarkError(result)


def MeasureCommand(command, iterations, env, track_memory):
    """Helper function to run arbitrary commands multiple times.

    Differences from MeasureGeneric():
        - MeasureGeneric() works with the performance/bm_*.py scripts.
        - MeasureCommand() does not echo every command run; it is intended for
          high-volume commands, like startup benchmarks

    Args:
        command: list of strings to be passed to Popen.
        iterations: number of times to run the command.
        env: environment vars dictionary.
        track_memory: bool to indicate whether to track memory usage.

    Returns:
        RawData instance. Note that we take instrumentation data from the final
        run; merging instrumentation data between multiple runs is
        prohibitively difficult at this point.

    Raises:
        RuntimeError: if the command failed.
    """
    with open(os.devnull, "wb") as dev_null:
        RemovePycs()

        # Priming run (create pyc files, etc).
        CallAndCaptureOutput(command, env=env)

        an_s = "s"
        if iterations == 1:
            an_s = ""
        info("Running `%s` %d time%s", command, iterations, an_s)

        times = []
        mem_usage = []
        for _ in range(iterations):
            start_time = GetChildUserTime()
            subproc = subprocess.Popen(command,
                                       stdout=dev_null, stderr=subprocess.PIPE,
                                       env=env)
            if track_memory:
                future = MemoryUsageFuture(subproc.pid)
            _, stderr = subproc.communicate()
            if subproc.returncode != 0:
                raise RuntimeError("Benchmark died: " + stderr)
            if track_memory:
                mem_samples = future.GetMemoryUsage()
            end_time = GetChildUserTime()
            elapsed = end_time - start_time
            assert elapsed != 0
            times.append(elapsed)
            if track_memory:
                mem_usage.extend(mem_samples)

    if not track_memory:
        mem_usage = None
    return RawData(times, mem_usage, inst_output=stderr)


def Measure2to3(python, options):
    fast_target = Relative("lib/2to3/lib2to3/refactor.py")
    two_to_three_bin = Relative("lib/2to3/2to3")
    two_to_three_dir = Relative("lib/2to3")
    env = BuildEnv({"PYTHONPATH": two_to_three_dir},
                   inherit_env=options.inherit_env)

    # This can be compressed, but it's harder to understand.
    if options.fast:
        trials = 1
        target = fast_target
    elif options.rigorous:
        trials = 5
        target = two_to_three_dir
    else:
        trials = 1
        target = two_to_three_dir

    command = python + [two_to_three_bin, "-f", "all", target]
    return MeasureCommand(command, trials, env, options.track_memory)


def BM_2to3(*args, **kwargs):
    return SimpleBenchmark(Measure2to3, *args, **kwargs)


def MeasureHgStartup(python, options):
    hg_bin = Relative("lib/mercurial/hg")
    hg_dir = Relative("lib/mercurial")
    hg_lib_dir = Relative("lib/mercurial/mercurial/pure")
    hg_path = os.pathsep.join([hg_dir, hg_lib_dir])
    hg_env = BuildEnv({"PYTHONPATH": hg_path}, options.inherit_env)

    trials = 500
    if options.rigorous:
        trials = 1000
    elif options.fast:
        trials = 100

    command = python + [hg_bin, "help"]
    return MeasureCommand(command, trials, hg_env, options.track_memory)


def BM_hg_startup(*args, **kwargs):
    return SimpleBenchmark(MeasureHgStartup, *args, **kwargs)


def MeasureBzrStartup(python, options):
    bzr_bin = Relative("lib/bazaar/bzr")
    bzr_path = Relative("lib/bazaar")
    bzr_env = BuildEnv({"PYTHONPATH": bzr_path}, options.inherit_env)

    trials = 100
    if options.rigorous:
        trials = 200
    elif options.fast:
        trials = 10

    command = python + [bzr_bin, "help"]
    return MeasureCommand(command, trials, bzr_env, options.track_memory)


def BM_bzr_startup(*args, **kwargs):
    return SimpleBenchmark(MeasureBzrStartup, *args, **kwargs)


DJANGO_DIR = Relative("lib/django")


def MeasureDjango(python, options):
    bm_path = Relative("performance/bm_django.py")
    bm_env = {"PYTHONPATH": DJANGO_DIR}
    return MeasureGeneric(python, options, bm_path, bm_env)


def BM_Django(*args, **kwargs):
    return SimpleBenchmark(MeasureDjango, *args, **kwargs)


def MeasureFloat(python, options):
    bm_path = Relative("performance/bm_float.py")
    return MeasureGeneric(python, options, bm_path, iteration_scaling=5)

def BM_Float(*args, **kwargs):
    return SimpleBenchmark(MeasureFloat, *args, **kwargs)


def MeasureRietveld(python, options):
    PYTHONPATH = os.pathsep.join([DJANGO_DIR,
				  # These paths are lifted from
				  # lib/google_appengine.appcfg.py.  Note that we use
				  # our own version of Django instead of Appengine's.
				  Relative("lib/google_appengine"),
				  Relative("lib/google_appengine/lib/antlr3"),
				  Relative("lib/google_appengine/lib/webob"),
				  Relative("lib/google_appengine/lib/yaml/lib"),
				  Relative("lib/rietveld")])
    bm_path = Relative("performance/bm_rietveld.py")
    bm_env = {"PYTHONPATH": PYTHONPATH, "DJANGO_SETTINGS_MODULE": "settings"}

    return MeasureGeneric(python, options, bm_path, bm_env)


def BM_Rietveld(*args, **kwargs):
    return SimpleBenchmark(MeasureRietveld, *args, **kwargs)


def _ComesWithPsyco(python):
    """Determine whether the given Python binary already has Psyco.

    If the answer is no, we should build it (see BuildPsyco()).

    Args:
        python: prefix of a command line for the Python binary.

    Returns:
        True if we can "import psyco" with the given Python, False if not.
    """
    try:
        with open(os.devnull, "wb") as dev_null:
            subprocess.check_call(python + ["-c", "import psyco"],
                                  stdout=dev_null, stderr=dev_null,
                                  env=BuildEnv())
        return True
    except subprocess.CalledProcessError:
        return False


def _BuildPsyco(python):
    """Build Psyco against the given Python binary.

    Args:
        python: prefix of a command line for the Python binary.

    Returns:
        Path to Psyco's build directory. Putting this on your PYTHONPATH will
        make "import psyco" work.
    """
    PSYCO_SRC_DIR = Relative("lib/psyco")

    info("Building Psyco for %s", python)
    psyco_build_dir = tempfile.mkdtemp()
    abs_python = os.path.abspath(python[0])
    with ChangeDir(PSYCO_SRC_DIR):
        subprocess.check_call(LogCall([abs_python, "setup.py", "build",
                                       "--build-lib=" + psyco_build_dir]))
    return psyco_build_dir


def MeasureSpitfire(python, options, env=None, extra_args=[]):
    """Use Spitfire to test a Python binary's performance.

    Args:
        python: prefix of a command line for the Python binary to test.
        options: optparse.Values instance.
        env: optional; dict of environment variables to pass to Python.
        extra_args: optional; list of arguments to append to the Python
            command.

    Returns:
        RawData instance.
    """
    bm_path = Relative("performance/bm_spitfire.py")
    return MeasureGeneric(python, options, bm_path, env, extra_args)


def MeasureSpitfireWithPsyco(python, options):
    """Use Spitfire to measure Python's performance.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.

    Returns:
        RawData instance.
    """
    SPITFIRE_DIR = Relative("lib/spitfire")

    psyco_dir = ""
    if not _ComesWithPsyco(python):
        psyco_dir = _BuildPsyco(python)

    env_dirs = filter(bool, [SPITFIRE_DIR, psyco_dir])
    spitfire_env = {"PYTHONPATH": os.pathsep.join(env_dirs)}

    try:
        return MeasureSpitfire(python, options, spitfire_env)
    finally:
        try:
            shutil.rmtree(psyco_dir)
        except OSError:
            pass


def BM_Spitfire(*args, **kwargs):
    return SimpleBenchmark(MeasureSpitfireWithPsyco, *args, **kwargs)


def BM_SlowSpitfire(base_python, changed_python, options):
    extra_args = ["--disable_psyco"]
    spitfire_env = {"PYTHONPATH": Relative("lib/spitfire")}

    try:
        changed_data = MeasureSpitfire(changed_python, options,
                                       spitfire_env, extra_args)
        base_data = MeasureSpitfire(base_python, options,
                                    spitfire_env, extra_args)
    except subprocess.CalledProcessError, e:
        return str(e)

    return CompareBenchmarkData(base_data, changed_data, options)


def MeasurePickle(python, options, extra_args):
    """Test the performance of Python's pickle implementations.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.
        extra_args: list of arguments to append to the command line.

    Returns:
        RawData instance.
    """
    bm_path = Relative("performance/bm_pickle.py")
    return MeasureGeneric(python, options, bm_path, extra_args=extra_args)


def _PickleBenchmark(base_python, changed_python, options, extra_args):
    """Test the performance of Python's pickle implementations.

    Args:
        base_python: prefix of a command line for the reference
                Python binary.
        changed_python: prefix of a command line for the
                experimental Python binary.
        options: optparse.Values instance.
        extra_args: list of arguments to append to the command line.

    Returns:
        Summary of whether the experiemental Python is better/worse than the
        baseline.
    """
    return SimpleBenchmark(MeasurePickle,
                           base_python, changed_python, options, extra_args)


def BM_Pickle(base_python, changed_python, options):
    args = ["--use_cpickle", "pickle"]
    return _PickleBenchmark(base_python, changed_python, options, args)

def BM_Unpickle(base_python, changed_python, options):
    args = ["--use_cpickle", "unpickle"]
    return _PickleBenchmark(base_python, changed_python, options, args)

def BM_Pickle_List(base_python, changed_python, options):
    args = ["--use_cpickle", "pickle_list"]
    return _PickleBenchmark(base_python, changed_python, options, args)

def BM_Unpickle_List(base_python, changed_python, options):
    args = ["--use_cpickle", "unpickle_list"]
    return _PickleBenchmark(base_python, changed_python, options, args)

def BM_Pickle_Dict(base_python, changed_python, options):
    args = ["--use_cpickle", "pickle_dict"]
    return _PickleBenchmark(base_python, changed_python, options, args)

def BM_SlowPickle(base_python, changed_python, options):
    return _PickleBenchmark(base_python, changed_python, options, ["pickle"])

def BM_SlowUnpickle(base_python, changed_python, options):
    return _PickleBenchmark(base_python, changed_python, options, ["unpickle"])


def MeasureNQueens(python, options):
    """Test the performance of an N-Queens solver.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.

    Returns:
        RawData instance.
    """
    bm_path = Relative("performance/bm_nqueens.py")
    return MeasureGeneric(python, options, bm_path)


def BM_NQueens(*args, **kwargs):
    return SimpleBenchmark(MeasureNQueens, *args, **kwargs)


def _StartupPython(command, mem_usage, track_memory, inherit_env):
    startup_env = BuildEnv(inherit_env=inherit_env)
    if not track_memory:
        subprocess.check_call(command, env=startup_env)
    else:
        subproc = subprocess.Popen(command, env=startup_env)
        future = MemoryUsageFuture(subproc.pid)
        if subproc.wait() != 0:
            raise RuntimeError("Startup benchmark died")
        mem_usage.extend(future.GetMemoryUsage())


def MeasureStartup(python, cmd_opts, num_loops, track_memory, inherit_env):
    times = []
    work = ""
    if track_memory:
        # Without this, Python may start and exit before the memory sampler
        # thread has time to work. We can't just do 'time.sleep(x)', because
        # under -S, 'import time' fails.
        work = "for _ in xrange(200000): pass"
    command = python + cmd_opts + ["-c", work]
    mem_usage = []
    info("Running `%s` %d times", command, num_loops * 20)
    for _ in xrange(num_loops):
        t0 = time.time()
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        _StartupPython(command, mem_usage, track_memory, inherit_env)
        t1 = time.time()
        times.append(t1 - t0)
    if not track_memory:
      mem_usage = None
    return RawData(times, mem_usage)


def BM_normal_startup(base_python, changed_python, options):
    if options.rigorous:
        num_loops = 100
    elif options.fast:
        num_loops = 5
    else:
        num_loops = 50

    opts = []
    changed_data = MeasureStartup(changed_python, opts, num_loops,
                                  options.track_memory, options.inherit_env)
    base_data = MeasureStartup(base_python, opts, num_loops,
                               options.track_memory, options.inherit_env)

    return CompareBenchmarkData(base_data, changed_data, options)


def BM_startup_nosite(base_python, changed_python, options):
    if options.rigorous:
        num_loops = 200
    elif options.fast:
        num_loops = 10
    else:
        num_loops = 100

    opts = ["-S"]
    changed_data = MeasureStartup(changed_python, opts, num_loops,
                                  options.track_memory, options.inherit_env)
    base_data = MeasureStartup(base_python, opts, num_loops,
                               options.track_memory, options.inherit_env)

    return CompareBenchmarkData(base_data, changed_data, options)


def MeasureRegexPerformance(python, options, bm_path):
    """Test the performance of Python's regex engine.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.
        bm_path: relative path; which benchmark script to run.

    Returns:
        RawData instance.
    """
    return MeasureGeneric(python, options, Relative(bm_path))


def RegexBenchmark(base_python, changed_python, options, bm_path):
    return SimpleBenchmark(MeasureRegexPerformance,
                           base_python, changed_python, options, bm_path)


def BM_regex_v8(base_python, changed_python, options):
    bm_path = "performance/bm_regex_v8.py"
    return RegexBenchmark(base_python, changed_python, options, bm_path)


def BM_regex_effbot(base_python, changed_python, options):
    bm_path = "performance/bm_regex_effbot.py"
    return RegexBenchmark(base_python, changed_python, options, bm_path)


def BM_regex_compile(base_python, changed_python, options):
    bm_path = "performance/bm_regex_compile.py"
    return RegexBenchmark(base_python, changed_python, options, bm_path)


def MeasureThreading(python, options, bm_name):
    """Test the performance of Python's threading support.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.
        bm_name: name of the threading benchmark to run.

    Returns:
        RawData instance.
    """
    bm_path = Relative("performance/bm_threading.py")
    return MeasureGeneric(python, options, bm_path, extra_args=[bm_name])


def ThreadingBenchmark(base_python, changed_python, options, bm_name):
    return SimpleBenchmark(MeasureThreading,
                           base_python, changed_python, options, bm_name)


def BM_threaded_count(base_python, changed_python, options):
    bm_name = "threaded_count"
    return ThreadingBenchmark(base_python, changed_python, options, bm_name)


def BM_iterative_count(base_python, changed_python, options):
    bm_name = "iterative_count"
    return ThreadingBenchmark(base_python, changed_python, options, bm_name)


def MeasureUnpackSequence(python, options):
    """Test the performance of sequence unpacking.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.

    Returns:
        RawData instance.
    """
    bm_path = Relative("performance/bm_unpack_sequence.py")
    return MeasureGeneric(python, options, bm_path, iteration_scaling=1000)


def BM_unpack_sequence(*args, **kwargs):
    return SimpleBenchmark(MeasureUnpackSequence, *args, **kwargs)


def MeasureCallSimple(python, options):
    bm_path = Relative("performance/bm_call_simple.py")
    return MeasureGeneric(python, options, bm_path, iteration_scaling=3)


def BM_call_simple(*args, **kwargs):
    return SimpleBenchmark(MeasureCallSimple, *args, **kwargs)


def MeasureCallMethod(python, options):
    bm_path = Relative("performance/bm_call_method.py")
    return MeasureGeneric(python, options, bm_path, iteration_scaling=3)


def BM_call_method(*args, **kwargs):
    return SimpleBenchmark(MeasureCallMethod, *args, **kwargs)


def MeasureCallMethodUnknown(python, options):
    bm_path = Relative("performance/bm_call_method_unknown.py")
    return MeasureGeneric(python, options, bm_path, iteration_scaling=3)


def BM_call_method_unknown(*args, **kwargs):
    return SimpleBenchmark(MeasureCallMethodUnknown, *args, **kwargs)


def MeasureCallMethodSlots(python, options):
    bm_path = Relative("performance/bm_call_method_slots.py")
    return MeasureGeneric(python, options, bm_path, iteration_scaling=3)


def BM_call_method_slots(*args, **kwargs):
    return SimpleBenchmark(MeasureCallMethodSlots, *args, **kwargs)


def MeasureNbody(python, options):
    """Test the performance of math operations using an n-body benchmark.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.

    Returns:
        RawData instance.
    """
    bm_path = Relative("performance/bm_nbody.py")
    return MeasureGeneric(python, options, bm_path)


def BM_nbody(*args, **kwargs):
    return SimpleBenchmark(MeasureNbody, *args, **kwargs)


def MeasureSpamBayes(python, options):
    """Test the performance of the SpamBayes spam filter and its tokenizer.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.

    Returns:
        RawData instance.
    """
    pypath = os.pathsep.join([Relative("lib/spambayes"), Relative("lib/lockfile")])
    bm_path = Relative("performance/bm_spambayes.py")
    bm_env = BuildEnv({"PYTHONPATH": pypath}, options.inherit_env)
    return MeasureGeneric(python, options, bm_path, bm_env)


def BM_spambayes(*args, **kwargs):
    return SimpleBenchmark(MeasureSpamBayes, *args, **kwargs)


def MeasureHtml5libWarmup(python, options):
    """Test the performance of the html5lib HTML 5 parser.

    Args:
        python: prefix of a command line for the Python binary.
        options: optparse.Values instance.

    Returns:
        RawData instance.
    """
    bm_path = Relative("performance/bm_html5lib.py")
    bm_env = BuildEnv({"PYTHONPATH": Relative("lib/html5lib")},
                      options.inherit_env)
    return MeasureGeneric(python, options, bm_path, bm_env,
                          iteration_scaling=0.10)


def BM_html5lib_warmup(*args, **kwargs):
    return SimpleBenchmark(MeasureHtml5libWarmup, *args, **kwargs)


def MeasureHtml5lib(python, options):
    bm_path = Relative("performance/bm_html5lib.py")
    bm_env = BuildEnv({"PYTHONPATH": Relative("lib/html5lib")},
                      options.inherit_env)

    trials = 5
    if options.fast:
        trials = 1
    elif options.rigorous:
        trials = 10

    command = python + [bm_path, "-n", "1"]
    return MeasureCommand(command, trials, bm_env, options.track_memory)


def BM_html5lib(*args, **kwargs):
    return SimpleBenchmark(MeasureHtml5lib, *args, **kwargs)


def MeasureRichards(python, options):
    bm_path = Relative("performance/bm_richards.py")
    return MeasureGeneric(python, options, bm_path)

def BM_richards(*args, **kwargs):
    return SimpleBenchmark(MeasureRichards, *args, **kwargs)

### End benchmarks, begin main entry point support.

def _FindAllBenchmarks(namespace):
    return dict((name[3:].lower(), func)
                for (name, func) in sorted(namespace.iteritems())
                if name.startswith("BM_"))

BENCH_FUNCS = _FindAllBenchmarks(globals())

# Benchmark groups. The "default" group is what's run if no -b option is
# specified.
# If you update the default group, be sure to update the module docstring, too.
# An "all" group which includes every benchmark perf.py knows about is generated
# automatically.
BENCH_GROUPS = {"default": ["2to3", "django", "nbody", "rietveld",
                            "slowspitfire", "slowpickle", "slowunpickle",
                            "spambayes"],
                "startup": ["normal_startup", "startup_nosite",
                            "bzr_startup", "hg_startup"],
                "regex": ["regex_v8", "regex_effbot", "regex_compile"],
                "threading": ["threaded_count", "iterative_count"],
                "cpickle": ["pickle", "unpickle"],
                "apps": ["2to3", "html5lib", "rietveld", "spambayes"],
                "calls": ["call_simple", "call_method", "call_method_slots",
                          "call_method_unknown"],
                "math": ["nbody", "float"],
               }


def _ExpandBenchmarkName(bm_name, bench_groups):
    """Recursively expand name benchmark names.

    Args:
        bm_name: string naming a benchmark or benchmark group.

    Yields:
        Names of actual benchmarks, with all group names fully expanded.
    """
    expansion = bench_groups.get(bm_name)
    if expansion:
        for name in expansion:
            for name in _ExpandBenchmarkName(name, bench_groups):
                yield name
    else:
        yield bm_name


def ParseBenchmarksOption(benchmarks_opt, bench_groups):
    """Parses and verifies the --benchmarks option.

    Args:
        benchmarks_opt: the string passed to the -b option on the command line.

    Returns:
        A set() of the names of the benchmarks to run.
    """
    legal_benchmarks = bench_groups["all"]
    benchmarks = benchmarks_opt.split(",")
    positive_benchmarks = set(
        bm.lower() for bm in benchmarks if bm and bm[0] != "-")
    negative_benchmarks = set(
        bm[1:].lower() for bm in benchmarks if bm and bm[0] == "-")

    should_run = set()
    if not positive_benchmarks:
        should_run = set(_ExpandBenchmarkName("default", bench_groups))

    for name in positive_benchmarks:
        for bm in _ExpandBenchmarkName(name, bench_groups):
            if bm not in legal_benchmarks:
                logging.warning("No benchmark named %s", bm)
            else:
                should_run.add(bm)
    for bm in negative_benchmarks:
        if bm in bench_groups:
            raise ValueError("Negative groups not supported: -%s" % bm)
        elif bm not in legal_benchmarks:
            logging.warning("No benchmark named %s", bm)
        else:
            should_run.remove(bm)
    return should_run


def ParsePythonArgsOption(python_args_opt):
    """Parses the --args option.

    Args:
        python_args_opt: the string passed to the -a option on the command line.

    Returns:
        A pair of lists: (base_python_args, changed_python_args).
    """
    args_pair = python_args_opt.split(",")
    base_args = args_pair[0].split()  # On whitespace.
    changed_args = base_args
    if len(args_pair) == 2:
        changed_args = args_pair[1].split()
    elif len(args_pair) > 2:
        logging.warning("Didn't expect two or more commas in --args flag: %s",
                        python_args_opt)
    return base_args, changed_args


def ParseEnvVars(option, opt_str, value, parser):
    """Parser callback to --inherit_env var names."""
    parser.values.inherit_env = [v for v in value.split(",") if v]


def ParseOutputStyle(option, opt_str, value, parser):
    if value not in ("normal", "table"):
        parser.error("Invalid output style: %r" % value)
    parser.values.output_style = value


def main(argv, bench_funcs=BENCH_FUNCS, bench_groups=BENCH_GROUPS):
    bench_groups = bench_groups.copy()
    all_benchmarks = bench_funcs.keys()
    bench_groups["all"] = all_benchmarks

    parser = optparse.OptionParser(
        usage="%prog [options] baseline_python changed_python",
        description=("Compares the performance of baseline_python with" +
                     " changed_python and prints a report."))
    parser.add_option("-r", "--rigorous", action="store_true",
                      help=("Spend longer running tests to get more" +
                            " accurate results"))
    parser.add_option("-f", "--fast", action="store_true",
                      help="Get rough answers quickly")
    parser.add_option("-v", "--verbose", action="store_true",
                      help="Print more output")
    parser.add_option("-m", "--track_memory", action="store_true",
                      help="Track memory usage. This only works on Linux.")
    parser.add_option("-a", "--args", default="",
                      help=("Pass extra arguments to the python binaries."
                            " If there is a comma in this option's value, the"
                            " arguments before the comma (interpreted as a"
                            " space-separated list) are passed to the baseline"
                            " python, and the arguments after are passed to the"
                            " changed python. If there's no comma, the same"
                            " options are passed to both."))
    parser.add_option("-b", "--benchmarks", metavar="BM_LIST", default="",
                      help=("Comma-separated list of benchmarks to run.  Can" +
                            " contain both positive and negative arguments:" +
                            "  --benchmarks=run_this,also_this,-not_this.  If" +
                            " there are no positive arguments, we'll run all" +
                            " benchmarks except the negative arguments. " +
                            " Otherwise we run only the positive arguments. " +
                            " Valid benchmarks are: " +
                            ", ".join(bench_groups.keys() + all_benchmarks)))
    parser.add_option("--inherit_env", metavar="VAR_LIST", type="string",
                      action="callback", callback=ParseEnvVars, default=[],
                      help=("Comma-separated list of environment variable names"
                            " that are inherited from the parent environment"
                            " when running benchmarking subprocesses."))
    parser.add_option("--disable_timelines", default=False, action="store_true",
                      help="Don't use Google charts for displaying timelines.")
    parser.add_option("--output_style", metavar="STYLE", type="string",
                      action="callback", callback=ParseOutputStyle,
                      default="normal",
                      help=("What style the benchmark output should take."
                            " Valid options are 'normal' and 'table'."
                            " Default is '%default'."))
    parser.add_option("--control_label", metavar="LABEL", type="string",
                      action="store", default="",
                      help="Optional label for the control binary")
    parser.add_option("--experiment_label", metavar="LABEL", type="string",
                      action="store", default="",
                      help="Optional label for the experiment binary")
    parser.add_option("--diff_instrumentation", action="store_true",
                      help=("Compare the --with-instrumentation output from two"
                            " Unladen Swallow binaries. This is useful for"
                            " examining many benchmarks for optimization"
                            " effects."))


    options, args = parser.parse_args(argv)
    if len(args) != 2:
        parser.error("incorrect number of arguments")
    base, changed = args
    options.base_binary = base
    options.changed_binary = changed

    if not options.control_label:
        options.control_label = options.base_binary
    if not options.experiment_label:
        options.experiment_label = options.changed_binary

    base_args, changed_args = ParsePythonArgsOption(options.args)
    base_cmd_prefix = [base] + base_args
    changed_cmd_prefix = [changed] + changed_args

    logging.basicConfig(level=logging.INFO)

    if options.track_memory:
        if CanGetMemoryUsage():
            info("Suppressing performance data due to --track_memory")
        else:
            # TODO(collinwinter): make this work on other platforms.
            parser.error("--track_memory requires Windows with PyWin32 or " +
                         "Linux 2.6.16 or above")

    if options.diff_instrumentation:
        info("Suppressing performance data due to --diff_instrumentation")

    should_run = ParseBenchmarksOption(options.benchmarks, bench_groups)

    results = []
    for name in sorted(should_run):
        func = bench_funcs[name]
        print "Running %s..." % name
        options.benchmark_name = name  # Easier than threading this everywhere.
        results.append((name, func(base_cmd_prefix, changed_cmd_prefix,
                                   options)))

    print
    print "Report on %s" % " ".join(platform.uname())
    if multiprocessing:
        print "Total CPU cores:", multiprocessing.cpu_count()
    if options.output_style == "normal":
        for name, result in results:
            print
            print "###", name, "###"
            print result
    elif options.output_style == "table":
        print FormatOutputAsTable(options.control_label,
                                  options.experiment_label,
                                  results)
    else:
        raise ValueError("Invalid output_style: %r" % options.output_style)
    return results

if __name__ == "__main__":
    main(sys.argv[1:])