foreign-platform newline support

[jackjansen]

BPO	476814
Nosy	@gvanrossum, @tim-one, @loewis, @warsaw, @jackjansen
Files	@univnewlines-20020414: Patch version 7 @univnewlines-doc-20020414: Docs v2 test_univnewlines.py: test script v5, 14-apr-02

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/warsaw'
closed_at = <Date 2002-05-22.20:30:16.000>
created_at = <Date 2001-10-31.16:41:38.000>
labels = ['interpreter-core']
title = 'foreign-platform newline support'
updated_at = <Date 2002-05-22.20:30:16.000>
user = 'https://github.com/jackjansen'

bugs.python.org fields:

activity = <Date 2002-05-22.20:30:16.000>
actor = 'barry'
assignee = 'barry'
closed = True
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2001-10-31.16:41:38.000>
creator = 'jackjansen'
dependencies = []
files = ['3725', '3726', '3727']
hgrepos = []
issue_num = 476814
keywords = ['patch']
message_count = 13.0
messages = ['38047', '38048', '38049', '38050', '38051', '38052', '38053', '38054', '38055', '38056', '38057', '38058', '38059']
nosy_count = 5.0
nosy_names = ['gvanrossum', 'tim.peters', 'loewis', 'barry', 'jackjansen']
pr_nums = []
priority = 'normal'
resolution = 'accepted'
stage = None
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue476814'
versions = []

[jackjansen]

This patch enables Python to interpret all known
newline conventions,
CR, LF or CRLF, on all platforms.

This support is enabled by configuring with
--with-universal-newlines
(so by default it is off, and everything should behave
as usual).

With universal newline support enabled two things
happen:

• When importing or otherwise parsing .py files any
  newline convention
  is accepted.
• Python code can pass a new "t" mode parameter to
  open() which
  reads files with any newline convention. "t" cannot
  be combined with
  any other mode flags like "w" or "+", for obvious
  reasons.

File objects have a new attribute "newlines" which
contains the type of
newlines encountered in the file (or None when no
newline has been seen,
or "mixed" if there were various types of newlines).

Also included is a test script which tests both file
I/O and parsing.

[gvanrossum]

Logged In: YES
user_id=6380

Tim, can you review this or pass it on to someone else who
has time?

Jack developed this patch after a discussion in which I was
involved in some of the design, but I won't have time to
look at it until December.

[tim-one]

Logged In: YES
user_id=31435

It would be better if get_line just called
Py_UniversalNewlineFgets (when appropriate) instead of
duplicating its logic inline.

Py_UniversalNewlineFgets and Py_UniversalNewlineFread
should deal with releasing the global lock themselves --
the correct granularity for lock release/reacquire is
around the C-level input routines (esp. for fread).

The new routines never check for I/O errors! Why not? It
seems essential.

The new Fgets checks for EOF at the end of the loop instead
of the top. This is surprising, and I stared a long time
in vain trying to guess why. Setting

newlinetypes |= NEWLINE_CR;

immediately after seeing an '\r' would be as fast (instead
of waiting to see EOF and then inferring the prior
existence of '\r' indirectly from the state of the
skipnextlf flag).

Speaking of which <wink>, the fobj tests in the inner loop
waste cycles. Set the local flag vrbls whether or not fobj
is NULL. When you're *out* of the inner loop you can
simply decline to store the new masks when fobj is NULL
(and you're already doing the latter anyway). A test and
branch inside the loop is much more expensive than or'ing
in a flag bit inside the loop, ditto harder to understand.

Floating the univ_newline test out of the loop (and
duplicating the loop body, one way for univ_newline true
and the other for it false) would also save a test and
branch on every character.

Doing fread one character at a time is very inefficient.
Since you know you need to obtain n characters in the end,
and that these transformations require reading at least n
characters, you could very profitably read n characters in
one gulp at the start, then switch to k at a time where k
is the number of \r\n pairs seen since the last fread
call. This is easier to code than it sounds <wink>.

It would be fine by me if you included (and initialized)
the new file-object fields all the time, whether or not
universal newlines are configured. I'd rather waste a few
bytes in a file object than see #ifdefs spread thru the
code.

I'll be damned if I can think of a quick way to do this
stuff on Windows -- native Windows fgets() is still the
only Windows handle we have on avoiding crushing thread
overhead inside MS's C library. I'll think some more about
it (the thrust still being to eliminate the 't' mode flag,
as whined about <wink> on Python-Dev).

[jackjansen]

Logged In: YES
user_id=45365

Here's a new version of the patch. To address your issues
one by one:

• get_line and Py_UniversalNewlineFgets are too difficult to
  integrate, at leat,
  I don't see how I could do it. The storage management of
  get_line gets in the way.

• The global lock comment I don't understand. The
  Universal... routines are
  replacements for fgets() and fread(), so have nothing to do
  with the interpreter lock.

• The logic of all three routines (get_line too) has changed
  and I've put comments in.
  I hope this addresses some of the points.

• If universal_newline is false for a certain PyFileObject
  we now immedeately take
  a quick exit via fgets() or fread().

There's also a new test script, that tests some more border
cases (like lines longer
than 100 characters, and a lone CR just before end of file).

[tim-one]

Logged In: YES
user_id=31435

Back to Jack -- and sorry for sitting on it so long.
Clearly this isn't making it into 2.2 in the core. As I
said on Python-Dev, I believe this needs a PEP: the design
decisions are debatable, so *should* be debated outside the
Mac community too. Note, though, that I can't stop you
from adding it to the 2.2 Mac distribution (if you want it
badly enough there).

If a PEP won't be written, I suggest finding someone else
to review it again; maybe Guido. Note that the patch needs
doc changes too. The patch to regrtest.py doesn't belong
here (I assume it just slipped in). There seems a lot of
code in support of the f_newlinetypes member, and the value
of that member isn't clear -- I can't imagine a good use
for it (maybe it's a Mac thing?). The implementation of
Py_UniversalNewlineFread appears incorrect to me: it reads
n bytes *every* time around the outer loop, no matter how
few characters are still required, and n doesn't change
inside the loop. The business about the GIL may be due to
the lack of docs: are, or are not, people supposed to
release the GIL themselves around calls to these guys?
It's not documented, and it appears your intent differed
from my guess. Finally, it would be better to call ferror
() after calling fread() instead of before it <wink>.

[jackjansen]

Logged In: YES
user_id=45365

This version of the patch addresses the bug in Py_UniversalNewlineFread and fixes up some minor details. Tim's other issues are addressed (at least: I think they are:-) in a forthcoming PEP.

[jackjansen]

Logged In: YES
user_id=45365

A new version of the patch. Main differences are that U is now the mode character to trigger universal newline input and --with-universal-newlines is default on.

[gvanrossum]

Logged In: YES
user_id=6380

Thanks! But there's no documentation. Could I twist your arm
for a separate doc patch?

I'm tempted to give this a +1, but I'd like to hear from MvL
and MAL to see if they foresee any interaction with their
PEP-262 implemetation.

[jackjansen]

Logged In: YES
user_id=45365

New doc patch, and new version of the patch that mainly allows the U to be specified (no-op) in non-univ-newline-builds.

[loewis]

Logged In: YES
user_id=21627

What is the rationale for making this a compile-time option?
It seems to complicate things, with no apparent advantage.

If this is for backwards compatibility, don't make it an
option: nobody will rebuild Python just to work around a
compatibility problem.

Apart from that, the patch looks goo.d

[jackjansen]

Logged In: YES
user_id=45365

A final tweaks: return a tuple of newline values in stead of 'mixed'.

[jackjansen]

Logged In: YES
user_id=45365

Barry,
I've checked in the code fixes, but I don't dare check in the documentation fixes myself, as I have no TeX and hence no way to test them. Could you do this, please, and also check that I'm following all the relevant guidelines?
Thanks!

[warsaw]

Logged In: YES
user_id=12800

I'll be committing the documentation changes momentarily.
You were pretty close, and fortunately I had Dr. Doco
sitting next to me to clean up a few nits.
I'm also going to check in a change to the builtin `file'
docstring.