bug#11894: 24.1.50; [PATCH] diff-apply-hunk can be off by 1 line when the hunk is 0-context pure removal

bug#11894: 24.1.50; [PATCH] diff-apply-hunk can be off by 1 line when the hunk is 0-context pure removal

[Dmitry Gutov]

[-- Attachment #1: Type: text/plain, Size: 1116 bytes --]

This applies both to context and unified diffs.

To get a 0-context hunk, you invoke diff command with -c0 or -U0 argument.

1. Example files (4 and 3 lines long):

test:
z
abc

def

test2:
z
abc
def

2. Run `diff -c0 test test2 > test.diff`:
*** test	2012-07-09 06:04:04.572209000 +0400
--- test2	2012-07-09 06:04:11.987150600 +0400
***************
*** 3 ****
-
--- 2 ----

3. Open test.diff in Emacs, then:
a) Do `C-u C-c C-a' (reverse hunk) -> see the empty line appear after 
"z", instead of after "abc".
b) Open test2, add empty line after "abc", go to test.diff window, do 
`C-c C-a' (apply hunk), see the newline between "z" and "abc" disappear 
instead.

Not sure if we can rely on the line number being always off by 1 in such 
hunks (there's no insertion, so, technically, the second line number in 
the header could be arbitrary), but at least 3 versions of diff across 2 
different OSes work the same in this regard.

Note that if you try to create such hunk with `diff-split-hunk' (by 
slicing it off a bigger hunk), the line number won't be off by 1.
Maybe that's a bug in `diff-split-hunk'.

--Dmitry

[-- Attachment #2: diff-mode.diff --]
[-- Type: text/plain, Size: 817 bytes --]

diff --git a/lisp/vc/diff-mode.el b/lisp/vc/diff-mode.el
index 9034ffe..e27b60f 100644
--- a/lisp/vc/diff-mode.el
+++ b/lisp/vc/diff-mode.el
@@ -1675,7 +1675,11 @@ NOPROMPT, if non-nil, means not to prompt the user."
       (when (> (prefix-numeric-value other-file) 8)
 	(setq diff-jump-to-old-file other))
       (with-current-buffer buf
-        (goto-char (point-min)) (forward-line (1- (string-to-number line)))
+        (let ((line-num (string-to-number line)))
+          ;; When the hunk is pure deletion, line number is off by 1.
+          (when (string= (if reverse (car old) (car new)) "")
+            (incf line-num))
+          (goto-char (point-min)) (forward-line (1- line-num)))
 	(let* ((orig-pos (point))
 	       (switched nil)
 	       ;; FIXME: Check for case where both OLD and NEW are found.

bug#11894: 24.1.50; [PATCH] diff-apply-hunk can be off by 1 line when the hunk is 0-context pure removal

[Stefan Monnier]

> Not sure if we can rely on the line number being always off by 1 in such
> hunks (there's no insertion, so, technically, the second line number in the
> header could be arbitrary), but at least 3 versions of diff across
> 2 different OSes work the same in this regard.

This looks like a bug in those versions of diff (not that I know
a non-buggy version, tho).  Could you report it to GNU diffutils.

> Note that if you try to create such hunk with `diff-split-hunk' (by slicing
> it off a bigger hunk), the line number won't be off by 1.

Oh, right, so there is a "version of diff" that doesn't have this bug ;-)

> Maybe that's a bug in `diff-split-hunk'.

I doubt it.  At least `patch' seems to agree with diff-mode.el:

   % diff -c0 footest1 footest2 |patch -o footest3 footest1
   patching file footest1
   patch: **** replacement text or line numbers mangled in hunk at line 8
   % 

> -        (goto-char (point-min)) (forward-line (1- (string-to-number line)))
> +        (let ((line-num (string-to-number line)))
> +          ;; When the hunk is pure deletion, line number is off by 1.
> +          (when (string= (if reverse (car old) (car new)) "")
> +            (incf line-num))
> +          (goto-char (point-min)) (forward-line (1- line-num)))

Context/unified diffs with 0 context are pretty rare, so I'd rather not
work around such a bug if I don't really have to.


        Stefan

bug#11894: 24.1.50; [PATCH] diff-apply-hunk can be off by 1 line when the hunk is 0-context pure removal

[Andreas Schwab]

Stefan Monnier <monnier@IRO.UMontreal.CA> writes:

> This looks like a bug in those versions of diff (not that I know
> a non-buggy version, tho).

Maybe, but that's how POSIX wants it:

    The ending line number of an empty range shall be the number of the
    preceding line, or 0 if the range is at the start of the file.

Andreas.

-- 
Andreas Schwab, schwab@linux-m68k.org
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

bug#11894: 24.1.50; [PATCH] diff-apply-hunk can be off by 1 line when the hunk is 0-context pure removal

[Stefan Monnier]

>> This looks like a bug in those versions of diff (not that I know
>> a non-buggy version, tho).
> Maybe, but that's how POSIX wants it:

>     The ending line number of an empty range shall be the number of the
>     preceding line, or 0 if the range is at the start of the file.

Hmm... but the number we're talking about is the starting-line-number,
isn't it?

In any case I think that we're OK in the sense that we're not doing
much worse than `patch'.


        Stefan

bug#11894: 24.1.50; [PATCH] diff-apply-hunk can be off by 1 line when the hunk is 0-context pure removal

[Andreas Schwab]

Stefan Monnier <monnier@IRO.UMontreal.CA> writes:

>>> This looks like a bug in those versions of diff (not that I know
>>> a non-buggy version, tho).
>> Maybe, but that's how POSIX wants it:
>
>>     The ending line number of an empty range shall be the number of the
>>     preceding line, or 0 if the range is at the start of the file.
>
> Hmm... but the number we're talking about is the starting-line-number,
> isn't it?

There is no starting-line-number.

    Next, the range of lines in file1 shall be written in the following
    format if the range contains two or more lines:

    "*** %d,%d ****\n", <beginning line number>, <ending line number>

    and the following format otherwise:

    "*** %d ****\n", <ending line number>

Andreas.

-- 
Andreas Schwab, schwab@linux-m68k.org
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

bug#11894: 24.1.50; [PATCH] diff-apply-hunk can be off by 1 line when the hunk is 0-context pure removal

[Dmitry Gutov]

On 18.07.2012 16:53, Stefan Monnier wrote:
>> Not sure if we can rely on the line number being always off by 1 in such
>> hunks (there's no insertion, so, technically, the second line number in the
>> header could be arbitrary), but at least 3 versions of diff across
>> 2 different OSes work the same in this regard.
>
> This looks like a bug in those versions of diff (not that I know
> a non-buggy version, tho).  Could you report it to GNU diffutils.

As per Andreas' comment, this is probably not a bug.

>> Note that if you try to create such hunk with `diff-split-hunk' (by slicing
>> it off a bigger hunk), the line number won't be off by 1.
>
> Oh, right, so there is a "version of diff" that doesn't have this bug ;-)

Yep, and I'll use it as a workaround if this bug won't be fixed.

>> Maybe that's a bug in `diff-split-hunk'.
>
> I doubt it.  At least `patch' seems to agree with diff-mode.el:
>
>     % diff -c0 footest1 footest2 |patch -o footest3 footest1
>     patching file footest1
>     patch: **** replacement text or line numbers mangled in hunk at line 8
>     %

`info diff` says: "For proper operation, `patch' typically needs at 
least two lines of context.", so this just may be a documented problem 
nobody cares about.

>> -        (goto-char (point-min)) (forward-line (1- (string-to-number line)))
>> +        (let ((line-num (string-to-number line)))
>> +          ;; When the hunk is pure deletion, line number is off by 1.
>> +          (when (string= (if reverse (car old) (car new)) "")
>> +            (incf line-num))
>> +          (goto-char (point-min)) (forward-line (1- line-num)))
>
> Context/unified diffs with 0 context are pretty rare, so I'd rather not
> work around such a bug if I don't really have to.

I'm fine either way, just wanted to report this.

--Dmitry