cpython/Tools/scripts/summarize_stats.py

"""Print a summary of specialization stats for all files in the
default stats folders.
"""

import argparse
import collections
import json
import os.path
import opcode
from datetime import date
import itertools
import sys

if os.name == "nt":
    DEFAULT_DIR = "c:\\temp\\py_stats\\"
else:
    DEFAULT_DIR = "/tmp/py_stats/"

#Create list of all instruction names
specialized = iter(opcode._specialized_instructions)
opname = ["<0>"]
for name in opcode.opname[1:]:
    if name.startswith("<"):
        try:
            name = next(specialized)
        except StopIteration:
            pass
    opname.append(name)

# opcode_name --> opcode
# Sort alphabetically.
opmap = {name: i for i, name in enumerate(opname)}
opmap = dict(sorted(opmap.items()))

TOTAL = "specialization.hit", "specialization.miss", "execution_count"

def format_ratio(num, den):
    """
    Format a ratio as a percentage. When the denominator is 0, returns the empty
    string.
    """
    if den == 0:
        return ""
    else:
        return f"{num/den:.01%}"

def join_rows(a_rows, b_rows):
    """
    Joins two tables together, side-by-side, where the first column in each is a
    common key.
    """
    if len(a_rows) == 0 and len(b_rows) == 0:
        return []

    if len(a_rows):
        a_ncols = list(set(len(x) for x in a_rows))
        if len(a_ncols) != 1:
            raise ValueError("Table a is ragged")

    if len(b_rows):
        b_ncols = list(set(len(x) for x in b_rows))
        if len(b_ncols) != 1:
            raise ValueError("Table b is ragged")

    if len(a_rows) and len(b_rows) and a_ncols[0] != b_ncols[0]:
        raise ValueError("Tables have different widths")

    if len(a_rows):
        ncols = a_ncols[0]
    else:
        ncols = b_ncols[0]

    default = [""] * (ncols - 1)
    a_data = {x[0]: x[1:] for x in a_rows}
    b_data = {x[0]: x[1:] for x in b_rows}

    if len(a_data) != len(a_rows) or len(b_data) != len(b_rows):
        raise ValueError("Duplicate keys")

    # To preserve ordering, use A's keys as is and then add any in B that aren't
    # in A
    keys = list(a_data.keys()) + [k for k in b_data.keys() if k not in a_data]
    return [(k, *a_data.get(k, default), *b_data.get(k, default)) for k in keys]

def calculate_specialization_stats(family_stats, total):
    rows = []
    for key in sorted(family_stats):
        if key.startswith("specialization.failure_kinds"):
            continue
        if key in ("specialization.hit", "specialization.miss"):
            label = key[len("specialization."):]
        elif key == "execution_count":
            continue
        elif key in ("specialization.success",  "specialization.failure", "specializable"):
            continue
        elif key.startswith("pair"):
            continue
        else:
            label = key
        rows.append((f"{label:>12}", f"{family_stats[key]:>12}", format_ratio(family_stats[key], total)))
    return rows

def calculate_specialization_success_failure(family_stats):
    total_attempts = 0
    for key in ("specialization.success",  "specialization.failure"):
        total_attempts += family_stats.get(key, 0)
    rows = []
    if total_attempts:
        for key in ("specialization.success",  "specialization.failure"):
            label = key[len("specialization."):]
            label = label[0].upper() + label[1:]
            val = family_stats.get(key, 0)
            rows.append((label, val, format_ratio(val, total_attempts)))
    return rows

def calculate_specialization_failure_kinds(name, family_stats, defines):
    total_failures = family_stats.get("specialization.failure", 0)
    failure_kinds = [ 0 ] * 40
    for key in family_stats:
        if not key.startswith("specialization.failure_kind"):
            continue
        _, index = key[:-1].split("[")
        index = int(index)
        failure_kinds[index] = family_stats[key]
    failures = [(value, index) for (index, value) in enumerate(failure_kinds)]
    failures.sort(reverse=True)
    rows = []
    for value, index in failures:
        if not value:
            continue
        rows.append((kind_to_text(index, defines, name), value, format_ratio(value, total_failures)))
    return rows

def print_specialization_stats(name, family_stats, defines):
    if "specializable" not in family_stats:
        return
    total = sum(family_stats.get(kind, 0) for kind in TOTAL)
    if total == 0:
        return
    with Section(name, 3, f"specialization stats for {name} family"):
        rows = calculate_specialization_stats(family_stats, total)
        emit_table(("Kind", "Count", "Ratio"), rows)
        rows = calculate_specialization_success_failure(family_stats)
        if rows:
            print_title("Specialization attempts", 4)
            emit_table(("", "Count:", "Ratio:"), rows)
            rows = calculate_specialization_failure_kinds(name, family_stats, defines)
            emit_table(("Failure kind", "Count:", "Ratio:"), rows)

def print_comparative_specialization_stats(name, base_family_stats, head_family_stats, defines):
    if "specializable" not in base_family_stats:
        return

    base_total = sum(base_family_stats.get(kind, 0) for kind in TOTAL)
    head_total = sum(head_family_stats.get(kind, 0) for kind in TOTAL)
    if base_total + head_total == 0:
        return
    with Section(name, 3, f"specialization stats for {name} family"):
        base_rows = calculate_specialization_stats(base_family_stats, base_total)
        head_rows = calculate_specialization_stats(head_family_stats, head_total)
        emit_table(
            ("Kind", "Base Count", "Base Ratio", "Head Count", "Head Ratio"),
            join_rows(base_rows, head_rows)
        )
        base_rows = calculate_specialization_success_failure(base_family_stats)
        head_rows = calculate_specialization_success_failure(head_family_stats)
        rows = join_rows(base_rows, head_rows)
        if rows:
            print_title("Specialization attempts", 4)
            emit_table(("", "Base Count:", "Base Ratio:", "Head Count:", "Head Ratio:"), rows)
            base_rows = calculate_specialization_failure_kinds(name, base_family_stats, defines)
            head_rows = calculate_specialization_failure_kinds(name, head_family_stats, defines)
            emit_table(
                ("Failure kind", "Base Count:", "Base Ratio:", "Head Count:", "Head Ratio:"),
                join_rows(base_rows, head_rows)
            )

def gather_stats(input):
    # Note the output of this function must be JSON-serializable

    if os.path.isfile(input):
        with open(input, "r") as fd:
            return json.load(fd)
    elif os.path.isdir(input):
        stats = collections.Counter()
        for filename in os.listdir(input):
            with open(os.path.join(input, filename)) as fd:
                for line in fd:
                    try:
                        key, value = line.split(":")
                    except ValueError:
                        print(f"Unparsable line: '{line.strip()}' in  {filename}", file=sys.stderr)
                        continue
                    key = key.strip()
                    value = int(value)
                    stats[key] += value
            stats['__nfiles__'] += 1
        return stats
    else:
        raise ValueError(f"{input:r} is not a file or directory path")

def extract_opcode_stats(stats):
    opcode_stats = [ {} for _ in range(256) ]
    for key, value in stats.items():
        if not key.startswith("opcode"):
            continue
        n, _, rest = key[7:].partition("]")
        opcode_stats[int(n)][rest.strip(".")] = value
    return opcode_stats

def parse_kinds(spec_src, prefix="SPEC_FAIL"):
    defines = collections.defaultdict(list)
    start = "#define " + prefix + "_"
    for line in spec_src:
        line = line.strip()
        if not line.startswith(start):
            continue
        line = line[len(start):]
        name, val = line.split()
        defines[int(val.strip())].append(name.strip())
    return defines

def pretty(defname):
    return defname.replace("_", " ").lower()

def kind_to_text(kind, defines, opname):
    if kind <= 8:
        return pretty(defines[kind][0])
    if opname == "LOAD_SUPER_ATTR":
        opname = "SUPER"
    elif opname.endswith("ATTR"):
        opname = "ATTR"
    elif opname in ("FOR_ITER", "SEND"):
        opname = "ITER"
    elif opname.endswith("SUBSCR"):
        opname = "SUBSCR"
    for name in defines[kind]:
        if name.startswith(opname):
            return pretty(name[len(opname)+1:])
    return "kind " + str(kind)

def categorized_counts(opcode_stats):
    basic = 0
    specialized = 0
    not_specialized = 0
    specialized_instructions = {
        op for op in opcode._specialized_instructions
        if "__" not in op}
    for i, opcode_stat in enumerate(opcode_stats):
        if "execution_count" not in opcode_stat:
            continue
        count = opcode_stat['execution_count']
        name = opname[i]
        if "specializable" in opcode_stat:
            not_specialized += count
        elif name in specialized_instructions:
            miss = opcode_stat.get("specialization.miss", 0)
            not_specialized += miss
            specialized += count - miss
        else:
            basic += count
    return basic, not_specialized, specialized

def print_title(name, level=2):
    print("#"*level, name)
    print()

class Section:

    def __init__(self, title, level=2, summary=None):
        self.title = title
        self.level = level
        if summary is None:
            self.summary = title.lower()
        else:
            self.summary = summary

    def __enter__(self):
        print_title(self.title, self.level)
        print("<details>")
        print("<summary>", self.summary, "</summary>")
        print()
        return self

    def __exit__(*args):
        print()
        print("</details>")
        print()

def to_str(x):
    if isinstance(x, int):
        return format(x, ",d")
    else:
        return str(x)

def emit_table(header, rows):
    width = len(header)
    header_line = "|"
    under_line = "|"
    for item in header:
        under = "---"
        if item.endswith(":"):
            item = item[:-1]
            under += ":"
        header_line += item + " | "
        under_line += under + "|"
    print(header_line)
    print(under_line)
    for row in rows:
        if width is not None and len(row) != width:
            raise ValueError("Wrong number of elements in row '" + str(row) + "'")
        print("|", " | ".join(to_str(i) for i in row), "|")
    print()

def calculate_execution_counts(opcode_stats, total):
    counts = []
    for i, opcode_stat in enumerate(opcode_stats):
        if "execution_count" in opcode_stat:
            count = opcode_stat['execution_count']
            miss = 0
            if "specializable" not in opcode_stat:
                miss = opcode_stat.get("specialization.miss")
            counts.append((count, opname[i], miss))
    counts.sort(reverse=True)
    cumulative = 0
    rows = []
    for (count, name, miss) in counts:
        cumulative += count
        if miss:
            miss = format_ratio(miss, count)
        else:
            miss = ""
        rows.append((name, count, format_ratio(count, total),
                     format_ratio(cumulative, total), miss))
    return rows

def emit_execution_counts(opcode_stats, total):
    with Section("Execution counts", summary="execution counts for all instructions"):
        rows = calculate_execution_counts(opcode_stats, total)
        emit_table(
            ("Name", "Count:", "Self:", "Cumulative:", "Miss ratio:"),
            rows
        )

def emit_comparative_execution_counts(
    base_opcode_stats, base_total, head_opcode_stats, head_total
):
    with Section("Execution counts", summary="execution counts for all instructions"):
        base_rows = calculate_execution_counts(base_opcode_stats, base_total)
        head_rows = calculate_execution_counts(head_opcode_stats, head_total)
        base_data = dict((x[0], x[1:]) for x in base_rows)
        head_data = dict((x[0], x[1:]) for x in head_rows)
        opcodes = set(base_data.keys()) | set(head_data.keys())

        rows = []
        default = [0, "0.0%", "0.0%", 0]
        for opcode in opcodes:
            base_entry = base_data.get(opcode, default)
            head_entry = head_data.get(opcode, default)
            if base_entry[0] == 0:
                change = 1
            else:
                change = (head_entry[0] - base_entry[0]) / base_entry[0]
            rows.append(
                (opcode, base_entry[0], head_entry[0],
                 f"{100*change:0.1f}%"))

        rows.sort(key=lambda x: -abs(float(x[-1][:-1])))

        emit_table(
            ("Name", "Base Count:", "Head Count:", "Change:"),
            rows
        )

def get_defines():
    spec_path = os.path.join(os.path.dirname(__file__), "../../Python/specialize.c")
    with open(spec_path) as spec_src:
        defines = parse_kinds(spec_src)
    return defines

def emit_specialization_stats(opcode_stats):
    defines = get_defines()
    with Section("Specialization stats", summary="specialization stats by family"):
        for i, opcode_stat in enumerate(opcode_stats):
            name = opname[i]
            print_specialization_stats(name, opcode_stat, defines)

def emit_comparative_specialization_stats(base_opcode_stats, head_opcode_stats):
    defines = get_defines()
    with Section("Specialization stats", summary="specialization stats by family"):
        for i, (base_opcode_stat, head_opcode_stat) in enumerate(zip(base_opcode_stats, head_opcode_stats)):
            name = opname[i]
            print_comparative_specialization_stats(name, base_opcode_stat, head_opcode_stat, defines)

def calculate_specialization_effectiveness(opcode_stats, total):
    basic, not_specialized, specialized = categorized_counts(opcode_stats)
    return [
        ("Basic", basic, format_ratio(basic, total)),
        ("Not specialized", not_specialized, format_ratio(not_specialized, total)),
        ("Specialized", specialized, format_ratio(specialized, total)),
    ]

def emit_specialization_overview(opcode_stats, total):
    with Section("Specialization effectiveness"):
        rows = calculate_specialization_effectiveness(opcode_stats, total)
        emit_table(("Instructions", "Count:", "Ratio:"), rows)
        for title, field in (("Deferred", "specialization.deferred"), ("Misses", "specialization.miss")):
            total = 0
            counts = []
            for i, opcode_stat in enumerate(opcode_stats):
                # Avoid double counting misses
                if title == "Misses" and "specializable" in opcode_stat:
                    continue
                value = opcode_stat.get(field, 0)
                counts.append((value, opname[i]))
                total += value
            counts.sort(reverse=True)
            if total:
                with Section(f"{title} by instruction", 3):
                    rows = [ (name, count, format_ratio(count, total)) for (count, name) in counts[:10] ]
                    emit_table(("Name", "Count:", "Ratio:"), rows)

def emit_comparative_specialization_overview(base_opcode_stats, base_total, head_opcode_stats, head_total):
    with Section("Specialization effectiveness"):
        base_rows = calculate_specialization_effectiveness(base_opcode_stats, base_total)
        head_rows = calculate_specialization_effectiveness(head_opcode_stats, head_total)
        emit_table(
            ("Instructions", "Base Count:", "Base Ratio:", "Head Count:", "Head Ratio:"),
            join_rows(base_rows, head_rows)
        )

def get_stats_defines():
    stats_path = os.path.join(os.path.dirname(__file__), "../../Include/pystats.h")
    with open(stats_path) as stats_src:
        defines = parse_kinds(stats_src, prefix="EVAL_CALL")
    return defines

def calculate_call_stats(stats):
    defines = get_stats_defines()
    total = 0
    for key, value in stats.items():
        if "Calls to" in key:
            total += value
            rows = []
    for key, value in stats.items():
        if "Calls to" in key:
            rows.append((key, value, format_ratio(value, total)))
        elif key.startswith("Calls "):
            name, index = key[:-1].split("[")
            index =  int(index)
            label = name + " (" + pretty(defines[index][0]) + ")"
            rows.append((label, value, format_ratio(value, total)))
    for key, value in stats.items():
        if key.startswith("Frame"):
            rows.append((key, value, format_ratio(value, total)))
    return rows

def emit_call_stats(stats):
    with Section("Call stats", summary="Inlined calls and frame stats"):
        rows = calculate_call_stats(stats)
        emit_table(("", "Count:", "Ratio:"), rows)

def emit_comparative_call_stats(base_stats, head_stats):
    with Section("Call stats", summary="Inlined calls and frame stats"):
        base_rows = calculate_call_stats(base_stats)
        head_rows = calculate_call_stats(head_stats)
        rows = join_rows(base_rows, head_rows)
        rows.sort(key=lambda x: -float(x[-1][:-1]))
        emit_table(
            ("", "Base Count:", "Base Ratio:", "Head Count:", "Head Ratio:"),
            rows
        )

def calculate_object_stats(stats):
    total_materializations = stats.get("Object new values")
    total_allocations = stats.get("Object allocations") + stats.get("Object allocations from freelist")
    total_increfs = stats.get("Object interpreter increfs") + stats.get("Object increfs")
    total_decrefs = stats.get("Object interpreter decrefs") + stats.get("Object decrefs")
    rows = []
    for key, value in stats.items():
        if key.startswith("Object"):
            if "materialize" in key:
                ratio = format_ratio(value, total_materializations)
            elif "allocations" in key:
                ratio = format_ratio(value, total_allocations)
            elif "increfs"     in key:
                ratio = format_ratio(value, total_increfs)
            elif "decrefs"     in key:
                ratio = format_ratio(value, total_decrefs)
            else:
                ratio = ""
            label = key[6:].strip()
            label = label[0].upper() + label[1:]
            rows.append((label, value, ratio))
    return rows

def emit_object_stats(stats):
    with Section("Object stats", summary="allocations, frees and dict materializatons"):
        rows = calculate_object_stats(stats)
        emit_table(("",  "Count:", "Ratio:"), rows)

def emit_comparative_object_stats(base_stats, head_stats):
    with Section("Object stats", summary="allocations, frees and dict materializatons"):
        base_rows = calculate_object_stats(base_stats)
        head_rows = calculate_object_stats(head_stats)
        emit_table(("",  "Base Count:", "Base Ratio:", "Head Count:", "Head Ratio:"), join_rows(base_rows, head_rows))

def get_total(opcode_stats):
    total = 0
    for opcode_stat in opcode_stats:
        if "execution_count" in opcode_stat:
            total += opcode_stat['execution_count']
    return total

def emit_pair_counts(opcode_stats, total):
    pair_counts = []
    for i, opcode_stat in enumerate(opcode_stats):
        if i == 0:
            continue
        for key, value in opcode_stat.items():
            if key.startswith("pair_count"):
                x, _, _ = key[11:].partition("]")
                if value:
                    pair_counts.append((value, (i, int(x))))
    with Section("Pair counts", summary="Pair counts for top 100 pairs"):
        pair_counts.sort(reverse=True)
        cumulative = 0
        rows = []
        for (count, pair) in itertools.islice(pair_counts, 100):
            i, j = pair
            cumulative += count
            rows.append((opname[i] + " " + opname[j], count, format_ratio(count, total),
                         format_ratio(cumulative, total)))
        emit_table(("Pair", "Count:", "Self:", "Cumulative:"),
            rows
        )
    with Section("Predecessor/Successor Pairs", summary="Top 5 predecessors and successors of each opcode"):
        predecessors = collections.defaultdict(collections.Counter)
        successors = collections.defaultdict(collections.Counter)
        total_predecessors = collections.Counter()
        total_successors = collections.Counter()
        for count, (first, second) in pair_counts:
            if count:
                predecessors[second][first] = count
                successors[first][second] = count
                total_predecessors[second] += count
                total_successors[first] += count
        for name, i in opmap.items():
            total1 = total_predecessors[i]
            total2 = total_successors[i]
            if total1 == 0 and total2 == 0:
                continue
            pred_rows = succ_rows = ()
            if total1:
                pred_rows = [(opname[pred], count, f"{count/total1:.1%}")
                             for (pred, count) in predecessors[i].most_common(5)]
            if total2:
                succ_rows = [(opname[succ], count, f"{count/total2:.1%}")
                             for (succ, count) in successors[i].most_common(5)]
            with Section(name, 3, f"Successors and predecessors for {name}"):
                emit_table(("Predecessors", "Count:", "Percentage:"),
                    pred_rows
                )
                emit_table(("Successors", "Count:", "Percentage:"),
                    succ_rows
                )

def output_single_stats(stats):
    opcode_stats = extract_opcode_stats(stats)
    total = get_total(opcode_stats)
    emit_execution_counts(opcode_stats, total)
    emit_pair_counts(opcode_stats, total)
    emit_specialization_stats(opcode_stats)
    emit_specialization_overview(opcode_stats, total)
    emit_call_stats(stats)
    emit_object_stats(stats)
    with Section("Meta stats", summary="Meta statistics"):
        emit_table(("", "Count:"), [('Number of data files', stats['__nfiles__'])])


def output_comparative_stats(base_stats, head_stats):
    base_opcode_stats = extract_opcode_stats(base_stats)
    base_total = get_total(base_opcode_stats)

    head_opcode_stats = extract_opcode_stats(head_stats)
    head_total = get_total(head_opcode_stats)

    emit_comparative_execution_counts(
        base_opcode_stats, base_total, head_opcode_stats, head_total
    )
    emit_comparative_specialization_stats(
        base_opcode_stats, head_opcode_stats
    )
    emit_comparative_specialization_overview(
        base_opcode_stats, base_total, head_opcode_stats, head_total
    )
    emit_comparative_call_stats(base_stats, head_stats)
    emit_comparative_object_stats(base_stats, head_stats)

def output_stats(inputs, json_output=None):
    if len(inputs) == 1:
        stats = gather_stats(inputs[0])
        if json_output is not None:
            json.dump(stats, json_output)
        output_single_stats(stats)
    elif len(inputs) == 2:
        if json_output is not None:
            raise ValueError(
                "Can not output to JSON when there are multiple inputs"
            )

        base_stats = gather_stats(inputs[0])
        head_stats = gather_stats(inputs[1])
        output_comparative_stats(base_stats, head_stats)

    print("---")
    print("Stats gathered on:", date.today())

def main():
    parser = argparse.ArgumentParser(description="Summarize pystats results")

    parser.add_argument(
        "inputs",
        nargs="*",
        type=str,
        default=[DEFAULT_DIR],
        help=f"""
        Input source(s).
        For each entry, if a .json file, the output provided by --json-output from a previous run;
        if a directory, a directory containing raw pystats .txt files.
        If one source is provided, its stats are printed.
        If two sources are provided, comparative stats are printed.
        Default is {DEFAULT_DIR}.
        """
    )

    parser.add_argument(
        "--json-output",
        nargs="?",
        type=argparse.FileType("w"),
        help="Output complete raw results to the given JSON file."
    )

    args = parser.parse_args()

    if len(args.inputs) > 2:
        raise ValueError("0-2 arguments may be provided.")

    output_stats(args.inputs, json_output=args.json_output)

if __name__ == "__main__":
    main()