mdiff: speed up showfunc for large diffs

This addresses the following issues with showfunc:

- Silly usage of regular expressions.
- Doing str.rstrip() needlessly in an inner loop.
- Doing catastrophic backtracking when trying to find a function line.

Finding function text is now at worst O(n lines in the old file), and
at best close to O(n hunks).

Given a diff like this[1]:

 src/main/antlr3/uk/ac/cam/ch/wwmm/pregenerated/ChemicalChunker.g        |      4 +-
 src/main/java/uk/ac/cam/ch/wwmm/pregenerated/ChemicalChunkerLexer.java  |      2 +-
 src/main/java/uk/ac/cam/ch/wwmm/pregenerated/ChemicalChunkerParser.java |  29189 +++++----
 3 files changed, 14741 insertions(+), 14454 deletions(-)

[1]: https://bitbucket.org/wwmm/chemicaltagger/changeset/d2bfbaecd4fc/raw

Without this change, hg log --stat --config diff.showfunc=1 takes an
absurdly long time to complete:

   CallCount    Recursive    Total(ms)   Inline(ms) module:lineno(function)
       32813            0     80.3546     40.6086   mercurial.mdiff:160(yieldhunk)
   [phone]            0     25.7227     25.7227   +<method 'match' of '_sre.SRE_Pattern' objects>
   [phone]            0     14.0221     14.0221   +<method 'rstrip' of 'str' objects>
       +1809            0      0.0009      0.0009   +mercurial.mdiff:148(contextend)
       +1809            0      0.0003      0.0003   +<len>
    65062746            0     25.7227     25.7227   <method 'match' of '_sre.SRE_Pattern' objects>
    65062763            0     14.0221     14.0221   <method 'rstrip' of 'str' objects>
         543            0      0.1631      0.1631   <zlib.decompress>
           3            0      0.0505      0.0505   <mercurial.bdiff.blocks>
       31007            0     80.4564      0.0477   mercurial.mdiff:147(_unidiff)
      +32813            0     80.3546     40.6086   +mercurial.mdiff:160(yieldhunk)
          +3            0      0.0505      0.0505   +<mercurial.bdiff.blocks>
       +3618            0      0.0022      0.0022   +mercurial.mdiff:154(contextstart)
       +5427            0      0.0013      0.0013   +<len>
          +3            0      0.0001      0.0000   +re:188(compile)
           1            0     80.8381      0.0322   mercurial.patch:1777(diffstatdata)
     +107499            0      0.0235      0.0235   +<method 'startswith' of 'str' objects>
      +31014            0     80.7820      0.0071   +mercurial.util:1284(iterlines)
          +3            0      0.0000      0.0000   +<method 'search' of '_sre.SRE_Pattern' objects>
          +4            0      0.0000      0.0000   +mercurial.patch:1783(addresult)
          +3            0      0.0000      0.0000   +<method 'group' of '_sre.SRE_Match' objects>
           6            0      0.0444      0.0283   mercurial.mdiff:12(splitnewlines)
          +6            0      0.0160      0.0160   +<method 'split' of 'str' objects>
          32            0      0.0246      0.0246   <method 'update' of '_hashlib.HASH' objects>
          11            0      0.0236      0.0236   <method 'read' of 'file' objects>
Time: real 80.880 secs (user 80.200+0.000 sys 0.380+0.000)

With this change, it's almost as fast as not using showfunc at all:

   CallCount    Recursive    Total(ms)   Inline(ms) module:lineno(function)
         543            0      0.1699      0.1699   <zlib.decompress>
           3            0      0.0501      0.0501   <mercurial.bdiff.blocks>
       32813            0      0.0415      0.0348   mercurial.mdiff:161(yieldhunk)
      +70837            0      0.0058      0.0058   +<method 'isalnum' of 'str' objects>
       +1809            0      0.0006      0.0006   +mercurial.mdiff:148(contextend)
       +1809            0      0.0002      0.0002   +<len>
           1            0      0.4879      0.0310   mercurial.patch:1777(diffstatdata)
     +107499            0      0.0230      0.0230   +<method 'startswith' of 'str' objects>
      +31014            0      0.4335      0.0065   +mercurial.util:1284(iterlines)
          +3            0      0.0000      0.0000   +<method 'search' of '_sre.SRE_Pattern' objects>
          +4            0      0.0000      0.0000   +mercurial.patch:1783(addresult)
          +1            0      0.0004      0.0000   +re:188(compile)
          32            0      0.0293      0.0293   <method 'update' of '_hashlib.HASH' objects>
           6            0      0.0427      0.0279   mercurial.mdiff:12(splitnewlines)
          +6            0      0.0147      0.0147   +<method 'split' of 'str' objects>
       31007            0      0.1169      0.0235   mercurial.mdiff:147(_unidiff)
          +3            0      0.0501      0.0501   +<mercurial.bdiff.blocks>
      +32813            0      0.0415      0.0348   +mercurial.mdiff:161(yieldhunk)
       +3618            0      0.0012      0.0012   +mercurial.mdiff:154(contextstart)
       +5427            0      0.0006      0.0006   +<len>
      107597            0      0.0230      0.0230   <method 'startswith' of 'str' objects>
          16            0      0.0213      0.0213   <mercurial.mpatch.patches>
         194            0      0.0149      0.0149   <method 'split' of 'str' objects>
Time: real 0.530 secs (user 0.450+0.000 sys 0.070+0.000)

mercurial/mdiff.py

@@ -1,277 +1,281 b''
 # mdiff.py - diff and patch routines for mercurial
 #
 # Copyright 2005, 2006 Matt Mackall <[e-mail]>
 #
 # This software may be used and distributed according to the terms of the
 # GNU General Public License version 2 or any later version.
 
 from i18n import _
 import bdiff, mpatch, util
 import re, struct
 
 def splitnewlines(text):
     '''like str.splitlines, but only split on newlines.'''
     lines = [l + '\n' for l in text.split('\n')]
     if lines:
         if lines[-1] == '\n':
             lines.pop()
         else:
             lines[-1] = lines[-1][:-1]
     return lines
 
 class diffopts(object):
     '''context is the number of context lines
     text treats all files as text
     showfunc enables diff -p output
     git enables the git extended patch format
     nodates removes dates from diff headers
     ignorews ignores all whitespace changes in the diff
     ignorewsamount ignores changes in the amount of whitespace
     ignoreblanklines ignores changes whose lines are all blank
     upgrade generates git diffs to avoid data loss
     '''
 
     defaults = {
         'context': 3,
         'text': False,
         'showfunc': False,
         'git': False,
         'nodates': False,
         'ignorews': False,
         'ignorewsamount': False,
         'ignoreblanklines': False,
         'upgrade': False,
         }
 
     __slots__ = defaults.keys()
 
     def __init__(self, **opts):
         for k in self.__slots__:
             v = opts.get(k)
             if v is None:
                 v = self.defaults[k]
             setattr(self, k, v)
 
         try:
             self.context = int(self.context)
         except ValueError:
             raise util.Abort(_('diff context lines count must be '
                                'an integer, not %r') % self.context)
 
     def copy(self, **kwargs):
         opts = dict((k, getattr(self, k)) for k in self.defaults)
         opts.update(kwargs)
         return diffopts(**opts)
 
 defaultopts = diffopts()
 
 def wsclean(opts, text, blank=True):
     if opts.ignorews:
         text = re.sub('[ \t\r]+', '', text)
     elif opts.ignorewsamount:
         text = re.sub('[ \t\r]+', ' ', text)
         text = text.replace(' \n', '\n')
     if blank and opts.ignoreblanklines:
         text = re.sub('\n+', '', text)
     return text
 
 def diffline(revs, a, b, opts):
     parts = ['diff']
     if opts.git:
         parts.append('--git')
     if revs and not opts.git:
         parts.append(' '.join(["-r %s" % rev for rev in revs]))
     if opts.git:
         parts.append('a/%s' % a)
         parts.append('b/%s' % b)
     else:
         parts.append(a)
     return ' '.join(parts) + '\n'
 
 def unidiff(a, ad, b, bd, fn1, fn2, r=None, opts=defaultopts):
     def datetag(date, addtab=True):
         if not opts.git and not opts.nodates:
             return '\t%s\n' % date
         if addtab and ' ' in fn1:
             return '\t\n'
         return '\n'
 
     if not a and not b:
         return ""
     epoch = util.datestr((0, 0))
 
     if not opts.text and (util.binary(a) or util.binary(b)):
         if a and b and len(a) == len(b) and a == b:
             return ""
         l = ['Binary file %s has changed\n' % fn1]
     elif not a:
         b = splitnewlines(b)
         if a is None:
             l1 = '--- /dev/null%s' % datetag(epoch, False)
         else:
             l1 = "--- %s%s" % ("a/" + fn1, datetag(ad))
         l2 = "+++ %s%s" % ("b/" + fn2, datetag(bd))
         l3 = "@@ -0,0 +1,%d @@\n" % len(b)
         l = [l1, l2, l3] + ["+" + e for e in b]
     elif not b:
         a = splitnewlines(a)
         l1 = "--- %s%s" % ("a/" + fn1, datetag(ad))
         if b is None:
             l2 = '+++ /dev/null%s' % datetag(epoch, False)
         else:
             l2 = "+++ %s%s" % ("b/" + fn2, datetag(bd))
         l3 = "@@ -1,%d +0,0 @@\n" % len(a)
         l = [l1, l2, l3] + ["-" + e for e in a]
     else:
         al = splitnewlines(a)
         bl = splitnewlines(b)
         l = list(_unidiff(a, b, al, bl, opts=opts))
         if not l:
             return ""
 
         l.insert(0, "--- a/%s%s" % (fn1, datetag(ad)))
         l.insert(1, "+++ b/%s%s" % (fn2, datetag(bd)))
 
     for ln in xrange(len(l)):
         if l[ln][-1] != '\n':
             l[ln] += "\n\ No newline at end of file\n"
 
     if r:
         l.insert(0, diffline(r, fn1, fn2, opts))
 
     return "".join(l)
 
 # creates a headerless unified diff
 # t1 and t2 are the text to be diffed
 # l1 and l2 are the text broken up into lines
 def _unidiff(t1, t2, l1, l2, opts=defaultopts):
     def contextend(l, len):
         ret = l + opts.context
         if ret > len:
             ret = len
         return ret
 
     def contextstart(l):
         ret = l - opts.context
         if ret < 0:
             return 0
         return ret
 
+    lastfunc = [0, '']
     def yieldhunk(hunk):
         (astart, a2, bstart, b2, delta) = hunk
         aend = contextend(a2, len(l1))
         alen = aend - astart
         blen = b2 - bstart + aend - a2
 
         func = ""
         if opts.showfunc:
-            # walk backwards from the start of the context
-            # to find a line starting with an alphanumeric char.
-            for x in xrange(astart - 1, -1, -1):
-                t = l1[x].rstrip()
-                if funcre.match(t):
-                    func = ' ' + t[:40]
+            lastpos, func = lastfunc
+            # walk backwards from the start of the context up to the start of
+            # the previous hunk context until we find a line starting with an
+            # alphanumeric char.
+            for i in xrange(astart - 1, lastpos - 1, -1):
+                if l1[i][0].isalnum():
+                    func = ' ' + l1[i].rstrip()[:40]
+                    lastfunc[1] = func
                     break
+            # by recording this hunk's starting point as the next place to
+            # start looking for function lines, we avoid reading any line in
+            # the file more than once.
+            lastfunc[0] = astart
 
         yield "@@ -%d,%d +%d,%d @@%s\n" % (astart + 1, alen,
                                            bstart + 1, blen, func)
         for x in delta:
             yield x
         for x in xrange(a2, aend):
             yield ' ' + l1[x]
 
-    if opts.showfunc:
-        funcre = re.compile('\w')
-
     # bdiff.blocks gives us the matching sequences in the files.  The loop
     # below finds the spaces between those matching sequences and translates
     # them into diff output.
     #
     if opts.ignorews or opts.ignorewsamount:
         t1 = wsclean(opts, t1, False)
         t2 = wsclean(opts, t2, False)
 
     diff = bdiff.blocks(t1, t2)
     hunk = None
     for i, s1 in enumerate(diff):
         # The first match is special.
         # we've either found a match starting at line 0 or a match later
         # in the file.  If it starts later, old and new below will both be
         # empty and we'll continue to the next match.
         if i > 0:
             s = diff[i - 1]
         else:
             s = [0, 0, 0, 0]
         delta = []
         a1 = s[1]
         a2 = s1[0]
         b1 = s[3]
         b2 = s1[2]
 
         old = l1[a1:a2]
         new = l2[b1:b2]
 
         # bdiff sometimes gives huge matches past eof, this check eats them,
         # and deals with the special first match case described above
         if not old and not new:
             continue
 
         if opts.ignoreblanklines:
             if wsclean(opts, "".join(old)) == wsclean(opts, "".join(new)):
                 continue
 
         astart = contextstart(a1)
         bstart = contextstart(b1)
         prev = None
         if hunk:
             # join with the previous hunk if it falls inside the context
             if astart < hunk[1] + opts.context + 1:
                 prev = hunk
                 astart = hunk[1]
                 bstart = hunk[3]
             else:
                 for x in yieldhunk(hunk):
                     yield x
         if prev:
             # we've joined the previous hunk, record the new ending points.
             hunk[1] = a2
             hunk[3] = b2
             delta = hunk[4]
         else:
             # create a new hunk
             hunk = [astart, a2, bstart, b2, delta]
 
         delta[len(delta):] = [' ' + x for x in l1[astart:a1]]
         delta[len(delta):] = ['-' + x for x in old]
         delta[len(delta):] = ['+' + x for x in new]
 
     if hunk:
         for x in yieldhunk(hunk):
             yield x
 
 def patchtext(bin):
     pos = 0
     t = []
     while pos < len(bin):
         p1, p2, l = struct.unpack(">lll", bin[pos:pos + 12])
         pos += 12
         t.append(bin[pos:pos + l])
         pos += l
     return "".join(t)
 
 def patch(a, bin):
     if len(a) == 0:
         # skip over trivial delta header
         return buffer(bin, 12)
     return mpatch.patches(a, [bin])
 
 # similar to difflib.SequenceMatcher.get_matching_blocks
 def get_matching_blocks(a, b):
     return [(d[0], d[2], d[1] - d[0]) for d in bdiff.blocks(a, b)]
 
 def trivialdiffheader(length):
     return struct.pack(">lll", 0, 0, length)
 
 patches = mpatch.patches
 patchedsize = mpatch.patchedsize
 textdiff = bdiff.bdiff