git diff annoyance / feature request

git diff annoyance / feature request

[Boaz Harrosh]

git diff has this very annoying miss-fixture where it will state
as hunk header the closest label instead of the function name.

So I get:
@@ -675,9 +670,23 @@ try_again:
 	}
 
 	if (flag) {
-		foo();
+		bazz();
 	}
 
 
Instead of what I'd like:
@@ -563,12 +563,7 @@ static int write_exec(struct page_collect *pcol)
 	}
 
 	if (flag) {
-		foo();
+		bazz();
 	}
 

I mean. The label "try_again" is not at all unique in my file. As a
reader I would like to see where is that code going to. The function
name is a unique file identifier that tells me exactly where the change
is going. The label is not. (It's not freaking BASIC)

I bet all this was just inherited from diff. Would it be accepted if
I send a patch to fix it? What you guys think a goto label makes any
sense at all?

Thanks
Boaz

Re: git diff annoyance / feature request

[Jeff King]

On Thu, Aug 25, 2011 at 12:14:24PM -0700, Boaz Harrosh wrote:

> git diff has this very annoying miss-fixture where it will state
> as hunk header the closest label instead of the function name.
> [...]
>
> I bet all this was just inherited from diff. Would it be accepted if
> I send a patch to fix it? What you guys think a goto label makes any
> sense at all?

Unless you tell git what type of content is in your file, it uses the same
basic heuristics for finding a hunk-header line that diff does. Namely,
the most recent line that starts with an alphabetic character,
underscore, or dollar sign.

If you want language-specific hunk headers, you can use gitattributes to
tell git what's in your files. We already have a builtin C driver that
will do what you want. You just need to do[1]:

  echo '*.c diff=cpp' >.gitattributes

Note that it handles both C and C++, hence the name. See "git help
gitattributes" for details (the section "Defining a custom hunk-header"
is what you want).

If your upstream (which looks like linux-2.6) doesn't want
.gitattributes files in the repository, you can also put the entry into
.git/info/attributes.

-Peff

[1] Since we have builtin funcname patterns for many types, we arguably
could also have a builtin mapping of common extensions to diff drivers.

Re: git diff annoyance / feature request

[Junio C Hamano]

Boaz Harrosh <bharrosh@panasas.com> writes:

> I mean. The label "try_again" is not at all unique in my file. As a
> reader I would like to see where is that code going to. The function
> name is a unique file identifier that tells me exactly where the change
> is going. The label is not. (It's not freaking BASIC)
>
> I bet all this was just inherited from diff. Would it be accepted if
> I send a patch to fix it? What you guys think a goto label makes any
> sense at all?

The default tries to mimic what GNU used to do when we added the feature.

The diff.*.xfuncname configuration variable is there exactly for people
like you to tweak what we use for hunk headers. Please experiment with it
and if you come up with a better set of patterns, people may want to copy
it and use it themselves. we may even consider updating the built-in
default with your patterns, once they got adopted by wider audiences.

Personally, I would have to say that the source wouldn't be using too many
labels with the same name for this behaviour to be problematic, especially
if it is not freaking BASIC ;-), so...

[RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Jeff King]

We already provide sane hunk-header patterns for specific
languages. However, the user has to manually map common
extensions to use them. It's not that hard to do, but it's
an extra step that the user might not even know is an
option. Let's be nice and do it automatically.

Signed-off-by: Jeff King <peff@peff.net>
---
I tried to think of negative side effects.

The userdiff drivers we have are pretty conservative; they just specify
hunk headers. So if you have a binary file named "foo.c", we still do
the regular binary detection.

If you have any matching attribute line in your own files, it should
override. So:

  foo/* -diff

will still mark foo/bar.c as binary, even with this change.

Can anyone think of other possible side effects?

Also, any other extensions that would go into such a list? I have no
idea what the common extension is for something like pascal or csharp.

 attr.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)

diff --git a/attr.c b/attr.c
index da29c8e..5118a14 100644
--- a/attr.c
+++ b/attr.c
@@ -294,6 +294,18 @@ static void free_attr_elem(struct attr_stack *e)
 
 static const char *builtin_attr[] = {
 	"[attr]binary -diff -text",
+	"*.html diff=html",
+	"*.java diff=java",
+	"*.perl diff=perl",
+	"*.pl diff=perl",
+	"*.php diff=php",
+	"*.py diff=python",
+	"*.rb diff=ruby",
+	"*.bib diff=bibtex",
+	"*.tex diff=tex",
+	"*.c diff=cpp",
+	"*.cc diff=cpp",
+	"*.cxx diff=cpp",
 	NULL,
 };
 
-- 
1.7.6.10.g62f04

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Eric Sunshine]

On Thu, Aug 25, 2011 at 4:40 PM, Jeff King <peff@peff.net> wrote:
> We already provide sane hunk-header patterns for specific
> languages. However, the user has to manually map common
> extensions to use them. It's not that hard to do, but it's
> an extra step that the user might not even know is an
> option. Let's be nice and do it automatically.
>
> Also, any other extensions that would go into such a list? I have no
> idea what the common extension is for something like pascal or csharp.

C# uses extension ".cs".

".cpp" is common, in fact often required, by Windows compilers.

What about ".h" and ".hpp"?

-- ES

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Jeff King]

On Thu, Aug 25, 2011 at 05:00:51PM -0400, Eric Sunshine wrote:

> > Also, any other extensions that would go into such a list? I have no
> > idea what the common extension is for something like pascal or csharp.
> 
> C# uses extension ".cs".
> 
> ".cpp" is common, in fact often required, by Windows compilers.

Thanks, added both to my list.

> What about ".h" and ".hpp"?

How well do our cpp patterns do with header files? I imagine they're
better than the default, but I don't think I've ever really tried
anything tricky.

-Peff

Re: git diff annoyance / feature request

[Boaz Harrosh]

On 08/25/2011 01:27 PM, Junio C Hamano wrote:
> Boaz Harrosh <bharrosh@panasas.com> writes:
> 
>> I mean. The label "try_again" is not at all unique in my file. As a
>> reader I would like to see where is that code going to. The function
>> name is a unique file identifier that tells me exactly where the change
>> is going. The label is not. (It's not freaking BASIC)
>>
>> I bet all this was just inherited from diff. Would it be accepted if
>> I send a patch to fix it? What you guys think a goto label makes any
>> sense at all?
> 
> The default tries to mimic what GNU used to do when we added the feature.
> 
> The diff.*.xfuncname configuration variable is there exactly for people
> like you to tweak what we use for hunk headers. Please experiment with it
> and if you come up with a better set of patterns, people may want to copy
> it and use it themselves. we may even consider updating the built-in
> default with your patterns, once they got adopted by wider audiences.
> 

Thanks, I'll investigate it sounds very interesting.

> Personally, I would have to say that the source wouldn't be using too many
> labels with the same name for this behaviour to be problematic, especially
> if it is not freaking BASIC ;-), so...

The Linux Kernel is full of "goto out" or "goto err" its a common error handling
practice. I actually like it because it taps onto a known pattern.

Now the patch tell me @@@ lable out: !! that's not very useful I would say

Thanks I'm sure I can shape it up the way I like it
Boaz

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Boaz Harrosh]

On 08/25/2011 02:06 PM, Jeff King wrote:
> On Thu, Aug 25, 2011 at 05:00:51PM -0400, Eric Sunshine wrote:
> 
>>> Also, any other extensions that would go into such a list? I have no
>>> idea what the common extension is for something like pascal or csharp.
>>
>> C# uses extension ".cs".
>>
>> ".cpp" is common, in fact often required, by Windows compilers.
> 
> Thanks, added both to my list.
> 
>> What about ".h" and ".hpp"?
> 
> How well do our cpp patterns do with header files? I imagine they're
> better than the default, but I don't think I've ever really tried
> anything tricky.
> 
> -Peff

Thanks Jeff, thanks everyone! This looks very promising. Specially that
it's all already there and I don't have to code it up.

RTFM time for me now
Boaz

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Brandon Casey]

On 08/25/2011 03:40 PM, Jeff King wrote:
> We already provide sane hunk-header patterns for specific
> languages. However, the user has to manually map common
> extensions to use them. It's not that hard to do, but it's
> an extra step that the user might not even know is an
> option. Let's be nice and do it automatically.
> 
> Signed-off-by: Jeff King <peff@peff.net>
> ---
> I tried to think of negative side effects.

That's what I worried about when I last touched this code.  Now I'm
thinking "what took us so long to do this!!??".

> Also, any other extensions that would go into such a list?

*.bib diff=bibtex
*.tex diff=tex

*.[Ff] diff=fortran
*.[Ff][0-9][0-9] diff=fortran

GNU fortran currently recognizes .fXX where XX is 90, 95, 03 and 08
and probably enables/disables features based on the respective standard.
[0-9][0-9] would future proof against fortran f13 and f25 as long as
there aren't other extensions that would conflict.

Wikipedia says that .for is an extension for fortran, but I've never
seen that in the wild.  Maybe it's a windows thing (3-char ext).

-Brandon

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Junio C Hamano]

Jeff King <peff@peff.net> writes:

> If you have any matching attribute line in your own files, it should
> override. So:
>
>   foo/* -diff
>
> will still mark foo/bar.c as binary, even with this change.
>
> Can anyone think of other possible side effects?
>
> Also, any other extensions that would go into such a list? I have no
> idea what the common extension is for something like pascal or csharp.

As long as the builtin ones are the lowest priority fallback, we should be
Ok.

Do we say anywhere that "Ah, this has 'diff' attribute defined, so it must
be text"? If so, we should fix _that_. In other words, having this one
extra entry

	"* diff=default"

in the builtin_attr[] array should be a no-op, I think.

>
>  attr.c |   12 ++++++++++++
>  1 files changed, 12 insertions(+), 0 deletions(-)
>
> diff --git a/attr.c b/attr.c
> index da29c8e..5118a14 100644
> --- a/attr.c
> +++ b/attr.c
> @@ -294,6 +294,18 @@ static void free_attr_elem(struct attr_stack *e)
>  
>  static const char *builtin_attr[] = {
>  	"[attr]binary -diff -text",
> +	"*.html diff=html",
> +	"*.java diff=java",
> +	"*.perl diff=perl",
> +	"*.pl diff=perl",
> +	"*.php diff=php",
> +	"*.py diff=python",
> +	"*.rb diff=ruby",
> +	"*.bib diff=bibtex",
> +	"*.tex diff=tex",
> +	"*.c diff=cpp",
> +	"*.cc diff=cpp",
> +	"*.cxx diff=cpp",
>  	NULL,
>  };

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Eric Sunshine]

On 08/25/2011 05:06 PM, Jeff King wrote:
> On Thu, Aug 25, 2011 at 05:00:51PM -0400, Eric Sunshine wrote:
>
>>> Also, any other extensions that would go into such a list? I have no
>>> idea what the common extension is for something like pascal or csharp.
>>
>> C# uses extension ".cs".
>>
>> ".cpp" is common, in fact often required, by Windows compilers.
>
> Thanks, added both to my list.

To clarify, I meant to say that for C++, .cpp is common/required on Windows.

>> What about ".h" and ".hpp"?
>
> How well do our cpp patterns do with header files? I imagine they're
> better than the default, but I don't think I've ever really tried
> anything tricky.

I scanned through a number of revisions for one of my long-running C++ 
projects comparing the diff of header files with and without "*.h 
diff=cpp". In some header files in this project, the oft-used C++ 
keywords public:, protected:, and private: appear at start-of-line. In 
such cases, the default diff emits a less-than-useful hunk header:

     @@ -19,8 +19,8 @@ public:

whereas, "diff=cpp" emits:

     @@ -19,8 +19,8 @@ class Foobar

-- ES

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Jeff King]

On Thu, Aug 25, 2011 at 07:44:25PM -0400, Eric Sunshine wrote:

> >How well do our cpp patterns do with header files? I imagine they're
> >better than the default, but I don't think I've ever really tried
> >anything tricky.
> 
> I scanned through a number of revisions for one of my long-running
> C++ projects comparing the diff of header files with and without "*.h
> diff=cpp". In some header files in this project, the oft-used C++
> keywords public:, protected:, and private: appear at start-of-line.
> In such cases, the default diff emits a less-than-useful hunk header:
> 
>     @@ -19,8 +19,8 @@ public:
> 
> whereas, "diff=cpp" emits:
> 
>     @@ -19,8 +19,8 @@ class Foobar

Thanks. My C++ is so rusty that I didn't think immediately of how often
those keywords appear in header files. Also, code in inline
functions in either C or C++ will be found in header files. So I think
defaulting *.h and *.hpp to cpp is sensible.

-Peff

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Jeff King]

On Thu, Aug 25, 2011 at 05:29:36PM -0500, Brandon Casey wrote:

> > Also, any other extensions that would go into such a list?
> 
> *.bib diff=bibtex
> *.tex diff=tex

I had those ones already. ;P

> *.[Ff] diff=fortran
> *.[Ff][0-9][0-9] diff=fortran

Thanks, I'll add those. I don't see a big problem with generalizing
f[0-9][0-9] to always be fortran, even though many of those numbers
aren't used. I don't think I've ever seen one used for anything else.

Should all of our matches be case-insensitive? That is, should we be
matching both .HTML and .html? Clearly lowercase is the One True Way,
but I don't know what kind of junk people with case-insensitive
filesystems have, or whether we should even worry about it.

> Wikipedia says that .for is an extension for fortran, but I've never
> seen that in the wild.  Maybe it's a windows thing (3-char ext).

We can leave it out. It's easy enough for somebody to add their own
gitattribute if they want, or to even complain that it should be in the
default set. I was trying to keep this list to the utterly common, not
become a catalogue of obscure fortran customs. :)

-Peff

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Jeff King]

On Thu, Aug 25, 2011 at 03:57:06PM -0700, Junio C Hamano wrote:

> > If you have any matching attribute line in your own files, it should
> > override. So:
> >
> >   foo/* -diff
> >
> > will still mark foo/bar.c as binary, even with this change.
> >
> > Can anyone think of other possible side effects?
> >
> > Also, any other extensions that would go into such a list? I have no
> > idea what the common extension is for something like pascal or csharp.
> 
> As long as the builtin ones are the lowest priority fallback, we should be
> Ok.
> 
> Do we say anywhere that "Ah, this has 'diff' attribute defined, so it must
> be text"? If so, we should fix _that_. In other words, having this one
> extra entry
> 
> 	"* diff=default"
> 
> in the builtin_attr[] array should be a no-op, I think.

No, certainly not since 122aa6f (diff: introduce diff.<driver>.binary,
2008-10-05). That commit's message claims that we did before it, but
looking at the patch, I am not so sure. But I'm not about to start
testing a 3-year-old patch to see if it really was the source of the
fix; the point is that it is correct now. :)

I think it could be a problem in the future if the builtin userdiff
drivers started growing more invasive options, like automatically
claiming to be non-binary (i.e., setting diff.cpp.binary = false by
default). In other words, I think we have two options:

  1. Builtin drivers like "cpp" can stay minimal, only setting funcname
     and color-words headers that aren't going to produce terrible
     results if we are wrong about detecting by extension.

  2. We force the user to identify file types manually, so we can't be
     wrong. The "cpp" diff driver means "you are a text C file", and if
     a user mis-marks a binary file with that diff driver, they are the
     one who is wrong.

So if it's an either/or situation, we should decide not only that
extension auto-detection is a good feature, but that it trumps adding
more advanced features to the builtin drivers in the future.

Or we could decide that the extensions really are good enough, and if
you really do have binary files named "foo.c", it's your problem to
override the defaults with "*.c -diff".

-Peff

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Eric Sunshine]

On 08/25/2011 10:45 PM, Jeff King wrote:
> Should all of our matches be case-insensitive? That is, should we be
> matching both .HTML and .html? Clearly lowercase is the One True Way,
> but I don't know what kind of junk people with case-insensitive
> filesystems have, or whether we should even worry about it.

In the Windows world, uppercase extensions are common. Also, one often 
finds .htm on Windows rather than .html.

Speaking of other platforms, on Mac OS X:

Objective-C is .m
Objective-C++ is .mm (and long-deprecated .M is probably not relevant)

-- ES

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Junio C Hamano]

Jeff King <peff@peff.net> writes:

> No, certainly not since 122aa6f (diff: introduce diff.<driver>.binary,
> 2008-10-05). That commit's message claims that we did before it, but
> looking at the patch, I am not so sure. But I'm not about to start
> testing a 3-year-old patch to see if it really was the source of the
> fix; the point is that it is correct now. :)

Violently agreed ;-)

> I think it could be a problem in the future if the builtin userdiff
> drivers started growing more invasive options, like automatically
> claiming to be non-binary (i.e., setting diff.cpp.binary = false by
> default).

Well, I think we can be careful when we start thinking about doing
something complex like that, then. Mentioning the above consideration in
the commit message of the final version of this patch would probably be a
good idea, I presume.

Thanks.

Re: git diff annoyance / feature request

[Miles Bader]

Boaz Harrosh <bharrosh@panasas.com> writes:
>> Personally, I would have to say that the source wouldn't be using too many
>> labels with the same name for this behaviour to be problematic, especially
>> if it is not freaking BASIC ;-), so...
>
> The Linux Kernel is full of "goto out" or "goto err" its a common error handling
> practice. I actually like it because it taps onto a known pattern.
>
> Now the patch tell me @@@ lable out: !! that's not very useful I would say
>
> Thanks I'm sure I can shape it up the way I like it

Incidentally, if these annoyances were inherited from GNU diff, it would
be good to send a bug report there too,... I doubt anybody likes the
behavior any better with diff!

[I know I've been annoyed by such things before...]

Thanks,

-Miles

-- 
Innards, n. pl. The stomach, heart, soul, and other bowels.

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Thomas Rast]

Jeff King wrote:
> We already provide sane hunk-header patterns for specific
> languages. However, the user has to manually map common
> extensions to use them. It's not that hard to do, but it's
> an extra step that the user might not even know is an
> option. Let's be nice and do it automatically.
> 
> Signed-off-by: Jeff King <peff@peff.net>
> ---
> I tried to think of negative side effects.
> 
> The userdiff drivers we have are pretty conservative; they just specify
> hunk headers.

And word-diff regexes.

In my book this is a plus for your patch, but I'm just saying.  It
will trigger the slightly slower mode of word-diffing that splits at
far more places than the default.

-- 
Thomas Rast
trast@{inf,student}.ethz.ch

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Brandon Casey]

On 08/25/2011 09:45 PM, Jeff King wrote:
> On Thu, Aug 25, 2011 at 05:29:36PM -0500, Brandon Casey wrote:
> 
>>> Also, any other extensions that would go into such a list?
>>
>> *.bib diff=bibtex
>> *.tex diff=tex
> 
> I had those ones already. ;P

Indeed.  I must be blind, I skipped right over them.

>> *.[Ff] diff=fortran
>> *.[Ff][0-9][0-9] diff=fortran
> 
> Thanks, I'll add those. I don't see a big problem with generalizing
> f[0-9][0-9] to always be fortran, even though many of those numbers
> aren't used. I don't think I've ever seen one used for anything else.
> 
> Should all of our matches be case-insensitive? That is, should we be
> matching both .HTML and .html? Clearly lowercase is the One True Way,
> but I don't know what kind of junk people with case-insensitive
> filesystems have, or whether we should even worry about it.

For the fortran case, Gnu fortran actually processes the files differently
depending on whether the f is capitalized (it preprocesses or not).  So
there is a functional reason for using a capital letter.

For the others, I don't know.  Do people still create files named .HTML
or is that just a relic of the past?  I can't really think of a strong
argument for or against matching insensitively.

-Brandon

Re: git diff annoyance / feature request

[René Scharfe]

Am 25.08.2011 21:14, schrieb Boaz Harrosh:
> 
> git diff has this very annoying miss-fixture where it will state
> as hunk header the closest label instead of the function name.
> 
> So I get:
> @@ -675,9 +670,23 @@ try_again:
>  	}
>  
>  	if (flag) {
> -		foo();
> +		bazz();
>  	}
>  
>  
> Instead of what I'd like:
> @@ -563,12 +563,7 @@ static int write_exec(struct page_collect *pcol)
>  	}
>  
>  	if (flag) {
> -		foo();
> +		bazz();
>  	}

Cheap trick: change your coding style to place a single space before
labels instead of having them start right at the beginning of a line.

René

Re: git diff annoyance / feature request

[Boaz Harrosh]

On 08/26/2011 02:16 PM, René Scharfe wrote:
> Am 25.08.2011 21:14, schrieb Boaz Harrosh:
> 
> Cheap trick: change your coding style to place a single space before
> labels instead of having them start right at the beginning of a line.
> 
> René
> 

Nope, does not work! and I have no choice about it, it's Linux coding
style

Boaz

Re: git diff annoyance / feature request

[Junio C Hamano]

Boaz Harrosh <bharrosh@panasas.com> writes:

> On 08/26/2011 02:16 PM, René Scharfe wrote:
>> Am 25.08.2011 21:14, schrieb Boaz Harrosh:
>> 
>> Cheap trick: change your coding style to place a single space before
>> labels instead of having them start right at the beginning of a line.
>> 
>> René
>> 
>
> Nope, does not work! and I have no choice about it, it's Linux coding
> style

I too thought it was in the Documentation/CodingStyle, but I don't find it
in my copy. "Chapter 7" has an example of using goto and it does have
label at the left edge of the page without indentation, but does not seem
to say it should not be indented.

Taken together with the output from "grep '#goto label' scripts/checkpatch.pl"
I concluded that "it's Linux coding style" was a myth.

Re: [RFC/PATCH] attr: map builtin userdiff drivers to well-known extensions

[Alexey Shumkin]

> Also, any other extensions that would go into such a list? I have no
> idea what the common extension is for something like pascal or csharp.

# [Object] Pascal unit files
*.pas diff=pascal

# + project files (they rarely contain procedures/functions
# but it is not forbidden in specification)

*.dpr diff=pascal

# for Lazarus - pp and .lpr respectivly
*.pp diff=pascal
*.lpr diff=pascal