Iterate over the lines of a string

Question

I have a multi-line string defined like this:

foo = """
this is 
a multi-line string.
"""

This string we used as test-input for a parser I am writing. The parser-function receives a file-object as input and iterates over it. It does also call the next() method directly to skip lines, so I really need an iterator as input, not an iterable. I need an iterator that iterates over the individual lines of that string like a file-object would over the lines of a text-file. I could of course do it like this:

lineiterator = iter(foo.splitlines())

Is there a more direct way of doing this? In this scenario the string has to traversed once for the splitting, and then again by the parser. It doesn't matter in my test-case, since the string is very short there, I am just asking out of curiosity. Python has so many useful and efficient built-ins for such stuff, but I could find nothing that suits this need.

Answer 1

Here are three possibilities:

foo = """
this is 
a multi-line string.
"""

def f1(foo=foo): return iter(foo.splitlines())

def f2(foo=foo):
    retval = ''
    for char in foo:
        retval += char if not char == '\n' else ''
        if char == '\n':
            yield retval
            retval = ''
    if retval:
        yield retval

def f3(foo=foo):
    prevnl = -1
    while True:
      nextnl = foo.find('\n', prevnl + 1)
      if nextnl < 0: break
      yield foo[prevnl + 1:nextnl]
      prevnl = nextnl

if __name__ == '__main__':
  for f in f1, f2, f3:
    print list(f())

Running this as the main script confirms the three functions are equivalent. With timeit (and a * 100 for foo to get substantial strings for more precise measurement):

$ python -mtimeit -s'import asp' 'list(asp.f3())'
1000 loops, best of 3: 370 usec per loop
$ python -mtimeit -s'import asp' 'list(asp.f2())'
1000 loops, best of 3: 1.36 msec per loop
$ python -mtimeit -s'import asp' 'list(asp.f1())'
10000 loops, best of 3: 61.5 usec per loop

Note we need the list() call to ensure the iterators are traversed, not just built.

IOW, the naive implementation is so much faster it isn't even funny: 6 times faster than my attempt with find calls, which in turn is 4 times faster than a lower-level approach.

Lessons to retain: measurement is always a good thing (but must be accurate); string methods like splitlines are implemented in very fast ways; putting strings together by programming at a very low level (esp. by loops of += of very small pieces) can be quite slow.

Edit: added @Jacob's proposal, slightly modified to give the same results as the others (trailing blanks on a line are kept), i.e.:

from cStringIO import StringIO

def f4(foo=foo):
    stri = StringIO(foo)
    while True:
        nl = stri.readline()
        if nl != '':
            yield nl.strip('\n')
        else:
            raise StopIteration

Measuring gives:

$ python -mtimeit -s'import asp' 'list(asp.f4())'
1000 loops, best of 3: 406 usec per loop

not quite as good as the .find based approach -- still, worth keeping in mind because it might be less prone to small off-by-one bugs (any loop where you see occurrences of +1 and -1, like my f3 above, should automatically trigger off-by-one suspicions -- and so should many loops which lack such tweaks and should have them -- though I believe my code is also right since I was able to check its output with other functions').

But the split-based approach still rules.

An aside: possibly better style for f4 would be:

from cStringIO import StringIO

def f4(foo=foo):
    stri = StringIO(foo)
    while True:
        nl = stri.readline()
        if nl == '': break
        yield nl.strip('\n')

at least, it's a bit less verbose. The need to strip trailing \ns unfortunately prohibits the clearer and faster replacement of the while loop with return iter(stri) (the iter part whereof is redundant in modern versions of Python, I believe since 2.3 or 2.4, but it's also innocuous). Maybe worth trying, also:

    return itertools.imap(lambda s: s.strip('\n'), stri)

or variations thereof -- but I'm stopping here since it's pretty much a theoretical exercise wrt the strip based, simplest and fastest, one.