Pytorch/Tools/scripts/fixdiv.py

#! /usr/bin/env python3

"""fixdiv - tool to fix division operators.

To use this tool, first run `python -Qwarnall yourscript.py 2>warnings'.
This runs the script `yourscript.py' while writing warning messages
about all uses of the classic division operator to the file
`warnings'.  The warnings look like this:

  <file>:<line>: DeprecationWarning: classic <type> division

The warnings are written to stderr, so you must use `2>' for the I/O
redirect.  I know of no way to redirect stderr on Windows in a DOS
box, so you will have to modify the script to set sys.stderr to some
kind of log file if you want to do this on Windows.

The warnings are not limited to the script; modules imported by the
script may also trigger warnings.  In fact a useful technique is to
write a test script specifically intended to exercise all code in a
particular module or set of modules.

Then run `python fixdiv.py warnings'.  This first reads the warnings,
looking for classic division warnings, and sorts them by file name and
line number.  Then, for each file that received at least one warning,
it parses the file and tries to match the warnings up to the division
operators found in the source code.  If it is successful, it writes
its findings to stdout, preceded by a line of dashes and a line of the
form:

  Index: <file>

If the only findings found are suggestions to change a / operator into
a // operator, the output is acceptable input for the Unix 'patch'
program.

Here are the possible messages on stdout (N stands for a line number):

- A plain-diff-style change ('NcN', a line marked by '<', a line
  containing '---', and a line marked by '>'):

  A / operator was found that should be changed to //.  This is the
  recommendation when only int and/or long arguments were seen.

- 'True division / operator at line N' and a line marked by '=':

  A / operator was found that can remain unchanged.  This is the
  recommendation when only float and/or complex arguments were seen.

- 'Ambiguous / operator (..., ...) at line N', line marked by '?':

  A / operator was found for which int or long as well as float or
  complex arguments were seen.  This is highly unlikely; if it occurs,
  you may have to restructure the code to keep the classic semantics,
  or maybe you don't care about the classic semantics.

- 'No conclusive evidence on line N', line marked by '*':

  A / operator was found for which no warnings were seen.  This could
  be code that was never executed, or code that was only executed
  with user-defined objects as arguments.  You will have to
  investigate further.  Note that // can be overloaded separately from
  /, using __floordiv__.  True division can also be separately
  overloaded, using __truediv__.  Classic division should be the same
  as either of those.  (XXX should I add a warning for division on
  user-defined objects, to disambiguate this case from code that was
  never executed?)

- 'Phantom ... warnings for line N', line marked by '*':

  A warning was seen for a line not containing a / operator.  The most
  likely cause is a warning about code executed by 'exec' or eval()
  (see note below), or an indirect invocation of the / operator, for
  example via the div() function in the operator module.  It could
  also be caused by a change to the file between the time the test
  script was run to collect warnings and the time fixdiv was run.

- 'More than one / operator in line N'; or
  'More than one / operator per statement in lines N-N':

  The scanner found more than one / operator on a single line, or in a
  statement split across multiple lines.  Because the warnings
  framework doesn't (and can't) show the offset within the line, and
  the code generator doesn't always give the correct line number for
  operations in a multi-line statement, we can't be sure whether all
  operators in the statement were executed.  To be on the safe side,
  by default a warning is issued about this case.  In practice, these
  cases are usually safe, and the -m option suppresses these warning.

- 'Can't find the / operator in line N', line marked by '*':

  This really shouldn't happen.  It means that the tokenize module
  reported a '/' operator but the line it returns didn't contain a '/'
  character at the indicated position.

- 'Bad warning for line N: XYZ', line marked by '*':

  This really shouldn't happen.  It means that a 'classic XYZ
  division' warning was read with XYZ being something other than
  'int', 'long', 'float', or 'complex'.

Notes:

- The augmented assignment operator /= is handled the same way as the
  / operator.

- This tool never looks at the // operator; no warnings are ever
  generated for use of this operator.

- This tool never looks at the / operator when a future division
  statement is in effect; no warnings are generated in this case, and
  because the tool only looks at files for which at least one classic
  division warning was seen, it will never look at files containing a
  future division statement.

- Warnings may be issued for code not read from a file, but executed
  using the exec() or eval() functions.  These may have
  <string> in the filename position, in which case the fixdiv script
  will attempt and fail to open a file named '<string>' and issue a
  warning about this failure; or these may be reported as 'Phantom'
  warnings (see above).  You're on your own to deal with these.  You
  could make all recommended changes and add a future division
  statement to all affected files, and then re-run the test script; it
  should not issue any warnings.  If there are any, and you have a
  hard time tracking down where they are generated, you can use the
  -Werror option to force an error instead of a first warning,
  generating a traceback.

- The tool should be run from the same directory as that from which
  the original script was run, otherwise it won't be able to open
  files given by relative pathnames.
"""

import sys
import getopt
import re
import tokenize

multi_ok = 0

def main():
    try:
        opts, args = getopt.getopt(sys.argv[1:], "hm")
    except getopt.error as msg:
        usage(msg)
        return 2
    for o, a in opts:
        if o == "-h":
            print(__doc__)
            return
        if o == "-m":
            global multi_ok
            multi_ok = 1
    if not args:
        usage("at least one file argument is required")
        return 2
    if args[1:]:
        sys.stderr.write("%s: extra file arguments ignored\n", sys.argv[0])
    warnings = readwarnings(args[0])
    if warnings is None:
        return 1
    files = list(warnings.keys())
    if not files:
        print("No classic division warnings read from", args[0])
        return
    files.sort()
    exit = None
    for filename in files:
        x = process(filename, warnings[filename])
        exit = exit or x
    return exit

def usage(msg):
    sys.stderr.write("%s: %s\n" % (sys.argv[0], msg))
    sys.stderr.write("Usage: %s [-m] warnings\n" % sys.argv[0])
    sys.stderr.write("Try `%s -h' for more information.\n" % sys.argv[0])

PATTERN = (r"^(.+?):(\d+): DeprecationWarning: "
           r"classic (int|long|float|complex) division$")

def readwarnings(warningsfile):
    prog = re.compile(PATTERN)
    warnings = {}
    try:
        f = open(warningsfile)
    except IOError as msg:
        sys.stderr.write("can't open: %s\n" % msg)
        return
    with f:
        while 1:
            line = f.readline()
            if not line:
                break
            m = prog.match(line)
            if not m:
                if line.find("division") >= 0:
                    sys.stderr.write("Warning: ignored input " + line)
                continue
            filename, lineno, what = m.groups()
            list = warnings.get(filename)
            if list is None:
                warnings[filename] = list = []
            list.append((int(lineno), sys.intern(what)))
    return warnings

def process(filename, list):
    print("-"*70)
    assert list # if this fails, readwarnings() is broken
    try:
        fp = open(filename)
    except IOError as msg:
        sys.stderr.write("can't open: %s\n" % msg)
        return 1
    with fp:
        print("Index:", filename)
        f = FileContext(fp)
        list.sort()
        index = 0 # list[:index] has been processed, list[index:] is still to do
        g = tokenize.generate_tokens(f.readline)
        while 1:
            startlineno, endlineno, slashes = lineinfo = scanline(g)
            if startlineno is None:
                break
            assert startlineno <= endlineno is not None
            orphans = []
            while index < len(list) and list[index][0] < startlineno:
                orphans.append(list[index])
                index += 1
            if orphans:
                reportphantomwarnings(orphans, f)
            warnings = []
            while index < len(list) and list[index][0] <= endlineno:
                warnings.append(list[index])
                index += 1
            if not slashes and not warnings:
                pass
            elif slashes and not warnings:
                report(slashes, "No conclusive evidence")
            elif warnings and not slashes:
                reportphantomwarnings(warnings, f)
            else:
                if len(slashes) > 1:
                    if not multi_ok:
                        rows = []
                        lastrow = None
                        for (row, col), line in slashes:
                            if row == lastrow:
                                continue
                            rows.append(row)
                            lastrow = row
                        assert rows
                        if len(rows) == 1:
                            print("*** More than one / operator in line", rows[0])
                        else:
                            print("*** More than one / operator per statement", end=' ')
                            print("in lines %d-%d" % (rows[0], rows[-1]))
                intlong = []
                floatcomplex = []
                bad = []
                for lineno, what in warnings:
                    if what in ("int", "long"):
                        intlong.append(what)
                    elif what in ("float", "complex"):
                        floatcomplex.append(what)
                    else:
                        bad.append(what)
                lastrow = None
                for (row, col), line in slashes:
                    if row == lastrow:
                        continue
                    lastrow = row
                    line = chop(line)
                    if line[col:col+1] != "/":
                        print("*** Can't find the / operator in line %d:" % row)
                        print("*", line)
                        continue
                    if bad:
                        print("*** Bad warning for line %d:" % row, bad)
                        print("*", line)
                    elif intlong and not floatcomplex:
                        print("%dc%d" % (row, row))
                        print("<", line)
                        print("---")
                        print(">", line[:col] + "/" + line[col:])
                    elif floatcomplex and not intlong:
                        print("True division / operator at line %d:" % row)
                        print("=", line)
                    elif intlong and floatcomplex:
                        print("*** Ambiguous / operator (%s, %s) at line %d:" %
                            ("|".join(intlong), "|".join(floatcomplex), row))
                        print("?", line)

def reportphantomwarnings(warnings, f):
    blocks = []
    lastrow = None
    lastblock = None
    for row, what in warnings:
        if row != lastrow:
            lastblock = [row]
            blocks.append(lastblock)
        lastblock.append(what)
    for block in blocks:
        row = block[0]
        whats = "/".join(block[1:])
        print("*** Phantom %s warnings for line %d:" % (whats, row))
        f.report(row, mark="*")

def report(slashes, message):
    lastrow = None
    for (row, col), line in slashes:
        if row != lastrow:
            print("*** %s on line %d:" % (message, row))
            print("*", chop(line))
            lastrow = row

class FileContext:
    def __init__(self, fp, window=5, lineno=1):
        self.fp = fp
        self.window = 5
        self.lineno = 1
        self.eoflookahead = 0
        self.lookahead = []
        self.buffer = []
    def fill(self):
        while len(self.lookahead) < self.window and not self.eoflookahead:
            line = self.fp.readline()
            if not line:
                self.eoflookahead = 1
                break
            self.lookahead.append(line)
    def readline(self):
        self.fill()
        if not self.lookahead:
            return ""
        line = self.lookahead.pop(0)
        self.buffer.append(line)
        self.lineno += 1
        return line
    def __getitem__(self, index):
        self.fill()
        bufstart = self.lineno - len(self.buffer)
        lookend = self.lineno + len(self.lookahead)
        if bufstart <= index < self.lineno:
            return self.buffer[index - bufstart]
        if self.lineno <= index < lookend:
            return self.lookahead[index - self.lineno]
        raise KeyError
    def report(self, first, last=None, mark="*"):
        if last is None:
            last = first
        for i in range(first, last+1):
            try:
                line = self[first]
            except KeyError:
                line = "<missing line>"
            print(mark, chop(line))

def scanline(g):
    slashes = []
    startlineno = None
    endlineno = None
    for type, token, start, end, line in g:
        endlineno = end[0]
        if startlineno is None:
            startlineno = endlineno
        if token in ("/", "/="):
            slashes.append((start, line))
        if type == tokenize.NEWLINE:
            break
    return startlineno, endlineno, slashes

def chop(line):
    if line.endswith("\n"):
        return line[:-1]
    else:
        return line

if __name__ == "__main__":
    sys.exit(main())