Fwd: Possibly nicer pathspec syntax?

Fwd: Possibly nicer pathspec syntax?

[Linus Torvalds]

[ Duh, I sent this just to Junio initially due to a brainfart. Here
goes the list also ]

Most of the time when I use pathspecs, I just use the bog-standard
normal ones, and everything works wonderfully.

But then occasionally I want to exclude a directory (usually
"drivers/"), and it all works fine, but the syntax for that is just
really cumbersome.

That's due to two issues:

 - the magic characters seem to have been chosen to be annoying on purpose

 - there's an extra "you can't exclude things without having positive
patterns" check

and both of those are rather nasty.

So to explain where I come from, during releases I do things like this:

    git diff -M --dirstat=2,cumulative v4.10-rc6..v4.10-rc7

and this is literaly why that "dirstat" diff exists - I find this very
useful for a project like the kernel that has a good hierarchical
directory structure. So the whole dirstat option came about from my
statistics gathering (see more explanations in my original commit
7df7c019c ("Add "--dirstat" for some directory statistics").

However, what often happens for the kernel is that a few big
subsystems end up dominating the discussion (usually drivers and
architecture), and then you want to drill down into everything else to
get that part. Long ago, that used to be painful, and I did things
like

    git diff -M --dirstat ... -- $(ls | egrep -v '(drivers)|(arch)')

which works, and gives me the dirstat for stuff that isn't arch or
driver updates.

However, git actually added exclude patterns, and I don't need to do
that crazy thing with shell expansion any more. Now I can do this
crazy thing with git natively instead:

    git diff -M --dirstat .. -- ':!drivers' ':!arch' .

but honestly, the git native interface really isn't much simpler than
what I used to do.

Is there really any reason for requiring the '.'?

[ Clarification from original message, since Junio asked: I didn't
actually want the semantics of '.' at all, since in a subdirectory it
limits to the current subdirectory. So I'd suggest that in the absense
of any positive pattern, there is simply no filtering at all, so
whenever I say '.' as a pattern, I really meant ":(top)." which is
even more of a cumbersom syntax that the current model really
encourages. Crazy. Since I tend to always work in the top directory,
the two are the same for me ]

And did we really have to pick such annoying characters that we need
the shell escaping?

(I never use the other ones with long forms, but they have the same
issue: parenthesis need escaping too, so you have to write them as

    ':(exclude,icase)drivers'

and you have to remember that a final colon is *not* allowed, and they
still need the escaping).

It really isn't all that wonderful to use from the command line.

In revisions, we use "^" for negation, partly exactly because '!' is
such a nasty character for shell users. With exclusion being the only
case I particularly use, I'd like that for pathspecs too, but maybe
others use icase etc?

IOW, what I'd like to do is just

    git diff -M --dirstat .. ^drivers ^arch

without needing the ugly quoting, and without needing the pointless
positive 'match everything else' marker.

Or even just allowing ^ in addition to ! for negation, and otherwise
keeping the current syntax.

[ Clarification from original message, since Junio asked: yes, this
suggestion still assumes the "don't need to specify the positive
pattern", so you could just do

    git diff :^drivers

  to avoid the drivers directory ]

Comments? Other ideas?

This is certainly not a high priority, but I hit it once again when
doing the 4.10-rc7 statistics, which is why I bring up the
discussion..

                 Linus

Re: Fwd: Possibly nicer pathspec syntax?

[Junio C Hamano]

Linus Torvalds <torvalds@linux-foundation.org> writes:

> [ Duh, I sent this just to Junio initially due to a brainfart. Here
> goes the list also ]

And my earlier response goes to the list ;-)

Linus Torvalds <torvalds@linux-foundation.org> writes:

> Most of the time when I use pathspecs, I just use the bog-standard
> normal ones, and everything works wonderfully.

It is, I think, a no-brainer to lift the "you must have at least one
positive".  If the user says "not this and that", it is reasonable
to assume that "but include everything else" is implied.

As to "!" that triggers history substitution without quoting, it may
be annoying and I think it is probably OK to pick a synonym letter,
perhaps "^", now that the set of pathspec magics do not seem to be
growing rapidly and there may not be any other existing magic that
the natural meaning of "^" would match better than "negate".  The
primary reason why we used ! is, I think, to match patterns in the
exclude files.

As to the leading ":", that is shared between the ":(long form)" and
the short form, I am a bit hesitant to lose it.  It allows the users
to be trained only once, i.e. "if you want to match a path without
magic in your working tree, you need to watch out for an unusual
path that begins with a colon, which may be quite minority to begin
with.  You just prefix ./ in front to defeat it.  Everything else
you can type as-is, modulo wildcard metacharacters, but you know
that already." and their brains need no upgrading.  Once we start
accepting short forms without the ":", every time we add a short
form magic, the users need to be retrained.

In short, this

> Or even just allowing ^ in addition to ! for negation, and otherwise
> keeping the current syntax.

in addition to "no positives?  let's pretend you also said '.' as a
positive", would not be too bad, methinks.  And that allows this

>     git diff -M --dirstat .. -- ':!drivers' ':!arch' .

to become

    git diff -M --dirstat .. -- :^{drivers,arch}

which is a bit shorter.  I personally am perfectly fine without ^, i.e.

    git diff -M --dirstat .. -- :\!{drivers,arch}

though.

By the way, I am wondering why this is private, not cc'ed to the
mailing list.  As messages addressed to gitster@ without git@vger
bypass my Inbox and gets thrown into spam box, which I only
occasionally scan to resurrect messages worth responding, and this
is one of those cases ;-)

Re: Fwd: Possibly nicer pathspec syntax?

[Linus Torvalds]

On Tue, 7 Feb 2017, Linus Torvalds wrote:
> 
> [ Clarification from original message, since Junio asked: I didn't
>   actually want the semantics of '.' at all, since in a subdirectory it
>   limits to the current subdirectory. So I'd suggest that in the absence
>   of any positive pattern, there is simply no filtering at all, so
>   whenever I say '.' as a pattern, I really meant ":(top)." which is
>   even more of a cumbersom syntax that the current model really
>   encourages. Crazy. Since I tend to always work in the top directory,
>   the two are the same for me ]

So here's an RFC patch, and I'm quoting the above part of my thinking 
because it's what the patch does, but it turns out that it's probably not 
what we want, and I suspect the "." behavior (as opposed to "no filtering 
at all") is actually better.

Now _I_ don't much care, since I only work from the top level, but without 
the "." behavior, you get into an odd situation that the negative match 
will be relative to the current directory, but then the positive matches 
will be everywhere else. 

Obviously, a negative match that has "top" set would change that logic. So 
this patch is purely a request for further discussion.

When I wrote the patch, I actually also removed the now stale entries from 
the 'po' files, but I'm not including that part here because it just 
distracts from the meat of it all. So this diff was actually generated 
with the new syntax:

	git diff -p --stat -- :^po/

and the only thing even remotely subtle here is that it changes our ctype 
array to make '^' be both a regex and a pathspec magic character.

Everything else should be pretty darn obvious.

The code *could* just track all the 'relative to top or not' bits in the 
exclusion pattern, and then use whatever top-ness the exclusion patterns 
have (and maybe fall back to the old warning if it had a mixture of 
exclusionary patterns). I'll happily change it to act that way if people 
think that makes sense.

Comments?

                Linus

---
 ctype.c    |  3 ++-
 pathspec.c | 15 ++++++++++-----
 2 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/ctype.c b/ctype.c
index fc0225ceb..250e2ce15 100644
--- a/ctype.c
+++ b/ctype.c
@@ -14,6 +14,7 @@ enum {
 	P = GIT_PATHSPEC_MAGIC, /* other non-alnum, except for ] and } */
 	X = GIT_CNTRL,
 	U = GIT_PUNCT,
+	Y = GIT_REGEX_SPECIAL | GIT_PATHSPEC_MAGIC,
 	Z = GIT_CNTRL | GIT_SPACE
 };
 
@@ -23,7 +24,7 @@ const unsigned char sane_ctype[256] = {
 	S, P, P, P, R, P, P, P, R, R, G, R, P, P, R, P,		/*  32.. 47 */
 	D, D, D, D, D, D, D, D, D, D, P, P, P, P, P, G,		/*  48.. 63 */
 	P, A, A, A, A, A, A, A, A, A, A, A, A, A, A, A,		/*  64.. 79 */
-	A, A, A, A, A, A, A, A, A, A, A, G, G, U, R, P,		/*  80.. 95 */
+	A, A, A, A, A, A, A, A, A, A, A, G, G, U, Y, P,		/*  80.. 95 */
 	P, A, A, A, A, A, A, A, A, A, A, A, A, A, A, A,		/*  96..111 */
 	A, A, A, A, A, A, A, A, A, A, A, R, R, U, P, X,		/* 112..127 */
 	/* Nothing in the 128.. range */
diff --git a/pathspec.c b/pathspec.c
index 7ababb315..ef59d080d 100644
--- a/pathspec.c
+++ b/pathspec.c
@@ -72,6 +72,7 @@ static struct pathspec_magic {
 	{ PATHSPEC_GLOB,    '\0', "glob" },
 	{ PATHSPEC_ICASE,   '\0', "icase" },
 	{ PATHSPEC_EXCLUDE,  '!', "exclude" },
+	{ PATHSPEC_EXCLUDE,  '^', "exclude" },
 };
 
 static void prefix_magic(struct strbuf *sb, int prefixlen, unsigned magic)
@@ -516,7 +517,7 @@ void parse_pathspec(struct pathspec *pathspec,
 	}
 
 	pathspec->nr = n;
-	ALLOC_ARRAY(pathspec->items, n);
+	ALLOC_ARRAY(pathspec->items, n+1);
 	item = pathspec->items;
 	prefixlen = prefix ? strlen(prefix) : 0;
 
@@ -540,10 +541,14 @@ void parse_pathspec(struct pathspec *pathspec,
 		pathspec->magic |= item[i].magic;
 	}
 
-	if (nr_exclude == n)
-		die(_("There is nothing to exclude from by :(exclude) patterns.\n"
-		      "Perhaps you forgot to add either ':/' or '.' ?"));
-
+	/*
+	 * If everything is an exclude pattern, add one positive pattern
+	 * that matches everyting. We allocated an extra one for this.
+	 */
+	if (nr_exclude == n) {
+		init_pathspec_item(item + n, 0, "", 0, "");
+		pathspec->nr++;
+	}
 
 	if (pathspec->magic & PATHSPEC_MAXDEPTH) {
 		if (flags & PATHSPEC_KEEP_ORDER)

Re: Fwd: Possibly nicer pathspec syntax?

[Mike Hommey]

On Tue, Feb 07, 2017 at 05:48:26PM -0800, Linus Torvalds wrote:
> 
> 
> On Tue, 7 Feb 2017, Linus Torvalds wrote:
> > 
> > [ Clarification from original message, since Junio asked: I didn't
> >   actually want the semantics of '.' at all, since in a subdirectory it
> >   limits to the current subdirectory. So I'd suggest that in the absence
> >   of any positive pattern, there is simply no filtering at all, so
> >   whenever I say '.' as a pattern, I really meant ":(top)." which is
> >   even more of a cumbersom syntax that the current model really
> >   encourages. Crazy. Since I tend to always work in the top directory,
> >   the two are the same for me ]
> 
> So here's an RFC patch, and I'm quoting the above part of my thinking 
> because it's what the patch does, but it turns out that it's probably not 
> what we want, and I suspect the "." behavior (as opposed to "no filtering 
> at all") is actually better.
> 
> Now _I_ don't much care, since I only work from the top level, but without 
> the "." behavior, you get into an odd situation that the negative match 
> will be relative to the current directory, but then the positive matches 
> will be everywhere else. 
> 
> Obviously, a negative match that has "top" set would change that logic. So 
> this patch is purely a request for further discussion.
> 
> When I wrote the patch, I actually also removed the now stale entries from 
> the 'po' files, but I'm not including that part here because it just 
> distracts from the meat of it all. So this diff was actually generated 
> with the new syntax:
> 
> 	git diff -p --stat -- :^po/
> 
> and the only thing even remotely subtle here is that it changes our ctype 
> array to make '^' be both a regex and a pathspec magic character.
> 
> Everything else should be pretty darn obvious.
> 
> The code *could* just track all the 'relative to top or not' bits in the 
> exclusion pattern, and then use whatever top-ness the exclusion patterns 
> have (and maybe fall back to the old warning if it had a mixture of 
> exclusionary patterns). I'll happily change it to act that way if people 
> think that makes sense.
> 
> Comments?

It seems to me that `git diff` and `git diff -- :^stuff` should have the
same output if there aren't changes in stuff, and `git diff` does the
same as `git diff -- :/` if you are in a subdirectory, not the same as
`git diff .`.

As such, the default positive match should be ':/' (which is shorter and
less cumbersome than ':(top)', btw)

Mike

Re: Fwd: Possibly nicer pathspec syntax?

[Junio C Hamano]

Linus Torvalds <torvalds@linux-foundation.org> writes:

> So here's an RFC patch, and I'm quoting the above part of my thinking 
> because it's what the patch does, but it turns out that it's probably not 
> what we want, and I suspect the "." behavior (as opposed to "no filtering 
> at all") is actually better.
> ...
>
> Comments?

 1. I think some commands limit their operands to cwd and some work
    on the whole tree when given no pathspec.  I think the "no
    positive?  then let's give you everything except these you
    excluded" should base the definition of "everything" to that.
    IOW, "cd t && git grep -e foo" shows everything in t/ directory,
    so the default you would add would be "." for "grep"; "cd t &&
    git diff HEAD~100 HEAD" would show everything, so you would give
    ":(top)." for "diff".

 2. I am not sure what ctype.c change is about.  Care to elaborate?

 3. I think our recent trend is to wean ourselves away from "an
    empty element in pathspec means all paths match", and I think we
    even have accepted a patch to emit a warning.  Doesn't the
    warning trigger for the new code below?

> -	if (nr_exclude == n)
> -		die(_("There is nothing to exclude from by :(exclude) patterns.\n"
> -		      "Perhaps you forgot to add either ':/' or '.' ?"));
> -
> +	/*
> +	 * If everything is an exclude pattern, add one positive pattern
> +	 * that matches everyting. We allocated an extra one for this.
> +	 */
> +	if (nr_exclude == n) {
> +		init_pathspec_item(item + n, 0, "", 0, "");
> +		pathspec->nr++;
> +	}
>  
>  	if (pathspec->magic & PATHSPEC_MAXDEPTH) {
>  		if (flags & PATHSPEC_KEEP_ORDER)

Re: Fwd: Possibly nicer pathspec syntax?

[Linus Torvalds]

On Tue, Feb 7, 2017 at 6:40 PM, Mike Hommey <mh@glandium.org> wrote:
>
> As such, the default positive match should be ':/' (which is shorter and
> less cumbersome than ':(top)', btw)

So that's what my patch does.

However, it's actually very counter-intuitive in a subdirectory.

Git doesn't do much of that, but let me give you an example from the
kernel. Again, this is not an example of anything I would do (because
I'm always at the top), but:

  [torvalds@i7 linux]$ cd drivers/
  [torvalds@i7 drivers]$ ll

  .. whee, *lots* of diorectories ..
  .. lets see what happened in net/ ..

  [torvalds@i7 drivers]$ git diff -M --dirstat=1,cumulative
v4.10-rc6..v4.10-rc7 -- net/
     7.4% drivers/net/ethernet/adaptec/
    47.9% drivers/net/ethernet/cadence/
     7.1% drivers/net/ethernet/emulex/benet/
     1.1% drivers/net/ethernet/freescale/
     3.6% drivers/net/ethernet/mellanox/mlx4/
    23.5% drivers/net/ethernet/mellanox/mlx5/core/
    27.2% drivers/net/ethernet/mellanox/
    92.5% drivers/net/ethernet/
     5.3% drivers/net/wireless/intel/iwlwifi/mvm/
     5.9% drivers/net/wireless/intel/iwlwifi/
   100.0% drivers/net/

  .. let's see what happened *outside* of net/ ..

[torvalds@i7 drivers]$ git diff -M --dirstat=1,cumulative
v4.10-rc6..v4.10-rc7 -- :^net/
   2.4% arch/arm64/crypto/
   2.1% arch/powerpc/include/asm/
   1.5% arch/powerpc/kernel/
   3.9% arch/powerpc/
   3.5% arch/sparc/kernel/
   4.1% arch/sparc/
   8.3% arch/x86/events/intel/
   1.7% arch/x86/kernel/cpu/mcheck/
   1.6% arch/x86/kernel/cpu/microcode/
   3.3% arch/x86/kernel/cpu/
   3.8% arch/x86/kernel/
   1.0% arch/x86/platform/efi/
  13.3% arch/x86/
  24.0% arch/
   1.1% drivers/base/
   2.9% drivers/dma/
  12.3% drivers/gpu/drm/i915/
   1.0% drivers/gpu/drm/nouveau/
  16.2% drivers/gpu/drm/
   3.9% drivers/hid/
   1.6% drivers/iio/
   2.3% drivers/regulator/
   ...

Notice? When you say "show only the net subdirectory" it does the
obvious thing relative to the current working directory, but if you
say "show everything _but_ the net subdirectory" it suddenly starts
showing other things.

Now, it would be easy enough to say "if you don't give a positive
path, we'll just use the empty path that matches the negative paths".
So if you ask for a negative relative "net" directory, we'll use the
relative empty path. And if you ask for a negative absolute path,
we'll use the empty absolute path.

It's a couple of lines more, and I think it might avoid some confusion.

And I suspect almost nobody has ever done any of this before,. because
the syntax was/is so cumbersome.

                Linus

Re: Fwd: Possibly nicer pathspec syntax?

[Linus Torvalds]

On Tue, Feb 7, 2017 at 6:42 PM, Junio C Hamano <gitster@pobox.com> wrote:
>
>  1. I think some commands limit their operands to cwd and some work
>     on the whole tree when given no pathspec. I think the "no
>     positive?  then let's give you everything except these you
>     excluded" should base the definition of "everything" to that.
>     IOW, "cd t && git grep -e foo" shows everything in t/ directory,
>     so the default you would add would be "." for "grep"; "cd t &&
>     git diff HEAD~100 HEAD" would show everything, so you would give
>     ":(top)." for "diff".

No. The thing is, "git diff" is relative too - for path
specifications. And the negative entries are pathspecs - and they act
as relative ones.

IOW, that whole

  cd drivers
  git diff A..B -- net/

will actually show the diff for drivers/net - so the pathspec very
much acts as relative to the cwd.

So no, absolute (ie ":(top)" or ":/") doesn't actually make sense for
"diff" either, even though diff by default is absolute when not given
a pathname at all.

But if you do

  cd drivers
  git diff A..B -- :^/arch

then suddenly an absolute positive root _does_ make sense,. because
now the negative pathspec was absolute..

Odd? Yes it is. But the positive pathspec rules are what they are, and
they are actually what I suspect everybody really wants. The existing
negative ones match the rules for the positive ones.

So I suspect that the best thing is if the "implicit positive rule
when there are no explicit ones" ends up matching the same semantics
as the (explicit) negative entries have..

>  2. I am not sure what ctype.c change is about.  Care to elaborate?

I didn't see the need for it either until I made the rest of the
patch, and it didn't work at all.

The pathspec.c code uses "if (is_pathspec_magic(..))" to test whether
a character is a short magiic pathspec character.  But '^' wasn't in
that set, because it was already marked as being (only) in the regex
set.

Does that whole is_pathspec_magic() thing make any sense when we have
an array that specifies the special characters we react to? No it does
not.

But it is what the code does, and I just made that code work.

>  3. I think our recent trend is to wean ourselves away from "an
>     empty element in pathspec means all paths match", and I think we
>     even have accepted a patch to emit a warning.  Doesn't the
>     warning trigger for the new code below?

It didn't trigger for me in my testing, I suspect the warning is at an
earlier level when it walks through the argv[] array and fills in the
pathspec arrays. But I didn't actually look for it.

                   Linus

Re: Fwd: Possibly nicer pathspec syntax?

[Mike Hommey]

On Tue, Feb 07, 2017 at 06:49:24PM -0800, Linus Torvalds wrote:
> On Tue, Feb 7, 2017 at 6:40 PM, Mike Hommey <mh@glandium.org> wrote:
> >
> > As such, the default positive match should be ':/' (which is shorter and
> > less cumbersome than ':(top)', btw)
> 
> So that's what my patch does.
> 
> However, it's actually very counter-intuitive in a subdirectory.
> 
> Git doesn't do much of that, but let me give you an example from the
> kernel. Again, this is not an example of anything I would do (because
> I'm always at the top), but:
> 
>   [torvalds@i7 linux]$ cd drivers/
>   [torvalds@i7 drivers]$ ll
> 
>   .. whee, *lots* of diorectories ..
>   .. lets see what happened in net/ ..
> 
>   [torvalds@i7 drivers]$ git diff -M --dirstat=1,cumulative
> v4.10-rc6..v4.10-rc7 -- net/
>      7.4% drivers/net/ethernet/adaptec/
>     47.9% drivers/net/ethernet/cadence/
>      7.1% drivers/net/ethernet/emulex/benet/
>      1.1% drivers/net/ethernet/freescale/
>      3.6% drivers/net/ethernet/mellanox/mlx4/
>     23.5% drivers/net/ethernet/mellanox/mlx5/core/
>     27.2% drivers/net/ethernet/mellanox/
>     92.5% drivers/net/ethernet/
>      5.3% drivers/net/wireless/intel/iwlwifi/mvm/
>      5.9% drivers/net/wireless/intel/iwlwifi/
>    100.0% drivers/net/
> 
>   .. let's see what happened *outside* of net/ ..
> 
> [torvalds@i7 drivers]$ git diff -M --dirstat=1,cumulative
> v4.10-rc6..v4.10-rc7 -- :^net/
>    2.4% arch/arm64/crypto/
>    2.1% arch/powerpc/include/asm/
>    1.5% arch/powerpc/kernel/
>    3.9% arch/powerpc/
>    3.5% arch/sparc/kernel/
>    4.1% arch/sparc/
>    8.3% arch/x86/events/intel/
>    1.7% arch/x86/kernel/cpu/mcheck/
>    1.6% arch/x86/kernel/cpu/microcode/
>    3.3% arch/x86/kernel/cpu/
>    3.8% arch/x86/kernel/
>    1.0% arch/x86/platform/efi/
>   13.3% arch/x86/
>   24.0% arch/
>    1.1% drivers/base/
>    2.9% drivers/dma/
>   12.3% drivers/gpu/drm/i915/
>    1.0% drivers/gpu/drm/nouveau/
>   16.2% drivers/gpu/drm/
>    3.9% drivers/hid/
>    1.6% drivers/iio/
>    2.3% drivers/regulator/
>    ...
> 
> Notice? When you say "show only the net subdirectory" it does the
> obvious thing relative to the current working directory, but if you
> say "show everything _but_ the net subdirectory" it suddenly starts
> showing other things.

I can totally see how this can be confusing, but the case can be made
that the problem is that `git diff` would be showing everything if you
didn't pass an exclusion list. Now, what you're suggesting introduces
some inconsistency, which, in itself, can cause confusion.

I'm not sure which confusion is worse.

Mike

Re: Fwd: Possibly nicer pathspec syntax?

[Junio C Hamano]

Linus Torvalds <torvalds@linux-foundation.org> writes:

> No. The thing is, "git diff" is relative too - for path
> specifications. And the negative entries are pathspecs - and they act
> as relative ones.
>
> IOW, that whole
>
>   cd drivers
>   git diff A..B -- net/
>
> will actually show the diff for drivers/net - so the pathspec very
> much acts as relative to the cwd.

But that is not what I was talking about.  Let's simplify.  I'd say
for any command that acts on "everything" when pathspec is not
given, the two sets of actual paths affected by these two:

	git cmd -- "net/"
	git cmd -- ":!net/"

should have no overlap (obviously) and when you take union of the
two sets, that should equal to

	git cmd --

i.e. no pathspecs.

>>  2. I am not sure what ctype.c change is about.  Care to elaborate?
>
> I didn't see the need for it either until I made the rest of the
> patch, and it didn't work at all.
>
> The pathspec.c code uses "if (is_pathspec_magic(..))" to test whether
> a character is a short magiic pathspec character.  But '^' wasn't in
> that set, because it was already marked as being (only) in the regex
> set.
>
> Does that whole is_pathspec_magic() thing make any sense when we have
> an array that specifies the special characters we react to? No it does
> not.
>
> But it is what the code does, and I just made that code work.

Ah, OK.

Re: Fwd: Possibly nicer pathspec syntax?

[Linus Torvalds]

[-- Attachment #1: Type: text/plain, Size: 2644 bytes --]

On Tue, Feb 7, 2017 at 7:12 PM, Junio C Hamano <gitster@pobox.com> wrote:
>
> But that is not what I was talking about.  Let's simplify.  I'd say
> for any command that acts on "everything" when pathspec is not
> given, the two sets of actual paths affected by these two:
>
>         git cmd -- "net/"
>         git cmd -- ":!net/"
>
> should have no overlap (obviously) and when you take union of the
> two sets, that should equal to
>
>         git cmd --
>
> i.e. no pathspecs.

Well, as mentioned, I won't ever care. I'm certainly ok with the "make
the default positive entry be everything".

I just suspect that from a user perspective that actually delves into
the subdirectories, the much bigger question will be: "I gave you a
pathspec, and suddenly you start giving me stuff from outside the area
entirely".

And then you can say "well, just add '.' to the pathspec", and you'd
be right, but I still think it's not what a naive user would expect.

People don't expect set theory from their pathspecs. They expect their
pathspecs to limit the output. They've learnt that within a
subdirectory, the pathspec limits to that subdirectory. And now it
suddenly starts showing things outside the subdirectory?

At that point no amount of "but but think about set theory" will
matter, methinks.

But I really don't feel strongly about it. The path I sent out (and
the slightly modified version attached in this email) actually acts
the way you suggest. It's certainly the simplest implementation. I
just suspect it's not the implementation people who go down into
subdirectories would want/expect.

>>>  2. I am not sure what ctype.c change is about.  Care to elaborate?
>>
>> I didn't see the need for it either until I made the rest of the
>> patch, and it didn't work at all.
>>
>> The pathspec.c code uses "if (is_pathspec_magic(..))" to test whether
>> a character is a short magiic pathspec character.  But '^' wasn't in
>> that set, because it was already marked as being (only) in the regex
>> set.
>>
>> Does that whole is_pathspec_magic() thing make any sense when we have
>> an array that specifies the special characters we react to? No it does
>> not.
>>
>> But it is what the code does, and I just made that code work.
>
> Ah, OK.

Side note: here's an alternative patch that avoids that issue
entirely, and also avoids a problem with prefix_magic() being uphappy
when one bit shows up multiple times in the array.

It's slightly hacky in parse_short_magic(), but it's smaller and
simpler, and avoids the two subtle cases. No need for ctype changes,
and no issues with prefix_magic() being somewhat stupid.

               Linus

[-- Attachment #2: patch.diff --]
[-- Type: text/plain, Size: 1293 bytes --]

 pathspec.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/pathspec.c b/pathspec.c
index 7ababb315..2a91247bc 100644
--- a/pathspec.c
+++ b/pathspec.c
@@ -224,6 +224,12 @@ static const char *parse_short_magic(unsigned *magic, const char *elem)
 		char ch = *pos;
 		int i;
 
+		// Special case alias for '!'
+		if (ch == '^') {
+			*magic |= PATHSPEC_EXCLUDE;
+			continue;
+		}
+
 		if (!is_pathspec_magic(ch))
 			break;
 
@@ -516,7 +522,7 @@ void parse_pathspec(struct pathspec *pathspec,
 	}
 
 	pathspec->nr = n;
-	ALLOC_ARRAY(pathspec->items, n);
+	ALLOC_ARRAY(pathspec->items, n+1);
 	item = pathspec->items;
 	prefixlen = prefix ? strlen(prefix) : 0;
 
@@ -540,10 +546,14 @@ void parse_pathspec(struct pathspec *pathspec,
 		pathspec->magic |= item[i].magic;
 	}
 
-	if (nr_exclude == n)
-		die(_("There is nothing to exclude from by :(exclude) patterns.\n"
-		      "Perhaps you forgot to add either ':/' or '.' ?"));
-
+	/*
+	 * If everything is an exclude pattern, add one positive pattern
+	 * that matches everyting. We allocated an extra one for this.
+	 */
+	if (nr_exclude == n) {
+		init_pathspec_item(item + n, 0, "", 0, "");
+		pathspec->nr++;
+	}
 
 	if (pathspec->magic & PATHSPEC_MAXDEPTH) {
 		if (flags & PATHSPEC_KEEP_ORDER)

Re: Fwd: Possibly nicer pathspec syntax?

[Junio C Hamano]

Linus Torvalds <torvalds@linux-foundation.org> writes:

> People don't expect set theory from their pathspecs. They expect their
> pathspecs to limit the output. They've learnt that within a
> subdirectory, the pathspec limits to that subdirectory. And now it
> suddenly starts showing things outside the subdirectory?
>
> At that point no amount of "but but think about set theory" will
> matter, methinks.

I do not feel too strongly about it, either, but when we invoke
"what would people who go down into subdirectories expect", the
issue becomes a lot bigger.

"git diff/log" without any pathspec in a subdirectory still shows
the whole diff.  "git grep" looks for hits inside that subdirectory,
on the other hand.

I think the former design decision is mostly a historical accident.
"The project tree as the whole is what matters" was one reason, and
another is that historically we didn't have ":/"--defaulting to the
whole tree and telling people to give "." was easier than defaulting
to the current and telling people to give "../.." after counting how
many levels deep you started at.  If we were designing the system
today with "." and ":/", I wouldn't be surprised if we chose "always
limit to cwd if there is no pathspec" for any command for the sake
of consistency.  We can easily say "if you want whole-tree, pass :/"
instead.

But we do not live in that world, and I do not think change them to
make things consistent is what we are discussing in this thread.
Given users accept and expect that "diff" without pathspec is whole
tree, while "grep" without pathspec is limited to cwd, what is the
reasonable definition of "everything from which negative ones are
excluded"?  That is the question I am trying to answer.

As Mike mentioned earlier in the thread, if we used "." (limit to
cwd) for "diff/log" when there are only negative pathspec elements,
the resulting UI would become inconsistent within a single command,
as in my world view, giving no pathspec means "work on everything
you would ordinarily work on", a positive pathspec means "among
everything you would ordinarily work on, only work on the ones that
match this pattern", and giving only a negative one ought to mean
"among everything you would ordinarily work on, only work on the
ones that do NOT match this pattern."

>  pathspec.c | 20 +++++++++++++++-----
>  1 file changed, 15 insertions(+), 5 deletions(-)
> 
> diff --git a/pathspec.c b/pathspec.c
> index 7ababb315..2a91247bc 100644
> --- a/pathspec.c
> +++ b/pathspec.c
> @@ -224,6 +224,12 @@ static const char *parse_short_magic(unsigned *magic, const char *elem)
>  		char ch = *pos;
>  		int i;
>  
> +		// Special case alias for '!'

/* style? */

> +		if (ch == '^') {
> +			*magic |= PATHSPEC_EXCLUDE;
> +			continue;
> +		}
> +
>  		if (!is_pathspec_magic(ch))
>  			break;
>  
> +	/*
> +	 * If everything is an exclude pattern, add one positive pattern
> +	 * that matches everyting. We allocated an extra one for this.
> +	 */
> +	if (nr_exclude == n) {
> +		init_pathspec_item(item + n, 0, "", 0, "");
> +		pathspec->nr++;
> +	}

I somehow do not think this is correct.  I expect

	cd t && git grep -e foo -- :^perf/

to look into things in 't' except for things in 't/perf', but the
above will grab hits from ../Documentation/ etc.  We need to pay
attention to PATHSPEC_PREFER_CWD bit in the flags word, I think.

A real change probably wants to become a two-patch series (one for
the new :^ alias, the other is for "everything except...") with log
message ;-)

Re: Fwd: Possibly nicer pathspec syntax?

[Linus Torvalds]

On Tue, 7 Feb 2017, Junio C Hamano wrote:
> > +		// Special case alias for '!'
> 
> /* style? */

Will fix.

> I somehow do not think this is correct.  I expect
> 
> 	cd t && git grep -e foo -- :^perf/
> 
> to look into things in 't' except for things in 't/perf', but the
> above will grab hits from ../Documentation/ etc.  We need to pay
> attention to PATHSPEC_PREFER_CWD bit in the flags word, I think.

Ok, that's easy enough.

Two-patch series to follow.

              Linus

Re: Fwd: Possibly nicer pathspec syntax?

[Duy Nguyen]

On Wed, Feb 8, 2017 at 12:12 PM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
> Two-patch series to follow.

glossary-content.txt update for both patches would be nice.
-- 
Duy

Re: Fwd: Possibly nicer pathspec syntax?

[Junio C Hamano]

Duy Nguyen <pclouds@gmail.com> writes:

> On Wed, Feb 8, 2017 at 12:12 PM, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
>> Two-patch series to follow.
>
> glossary-content.txt update for both patches would be nice.

I am no longer worried about it as I saw somebody actually sent
follow-up patches on this, but I want to pick your brain on one
thing that is related to this codepath.

We have PATHSPEC_PREFER_CWD and PATHSPEC_PREFER_FULL bits in flags,
added at fc12261fea ("parse_pathspec: add PATHSPEC_PREFER_{CWD,FULL}
flags", 2013-07-14), and I think the intent is some commands when
given no pathspec work on all paths in the current subdirectory
while others work on the full tree, regardless of where you are.
"grep" is in the former camp, "log" is in the latter.  And there is
a check to catch a bug in a caller that sets both.

I am wondering about this hunk (this is from the original commit
that added it):

 	if (!entry) {
 		static const char *raw[2];
 
+		if (flags & PATHSPEC_PREFER_FULL)
+			return;
+
+		if (!(flags & PATHSPEC_PREFER_CWD))
+			die("BUG: PATHSPEC_PREFER_CWD requires arguments");
+
 		pathspec->items = item = xmalloc(sizeof(*item));
 		memset(item, 0, sizeof(*item));
 		item->match = prefix;
		... returns a single entry pathspec to cover cwd ...

The BUG message is given when 

 - The command got no pathspec from the caller; and
 - PATHSPEC_PREFER_FULL is not set; and
 - PATHSPEC_PREFER_CWD is NOT set.

but the message says that the caller must have args when it sets
prefer-cwd.  Is this a simple typo?  If so what should it say?

	die("BUG: one of PATHSPEC_PREFER_FULL or _CWD must be set");

If that were the case, we are expressing only one bit of information
(do we limit to cwd, or do we work on full-tree?), but there must
have been a reason why we added two bits and made them mutually
incompatible so that we can express three possibilities.

Does this third possibility (i.e. a caller is allowed to pass
"flags" that does not prefer either) exist to support a command
where the caller MUST have at least one pathspec?  If that were the
case, this wouldn't be a BUG but an end-user error, e.g.

	die("at least one pathspec element is required");

If you know offhand which callers pass neither of the two
PATHSPEC_PREFER_* bits and remember for what purpose you allowed
them to do so, please remind me.  I'll keep digging in the meantime.

Thanks.

Re: Fwd: Possibly nicer pathspec syntax?

[Junio C Hamano]

Junio C Hamano <gitster@pobox.com> writes:

> If you know offhand which callers pass neither of the two
> PATHSPEC_PREFER_* bits and remember for what purpose you allowed
> them to do so, please remind me.  I'll keep digging in the meantime.

Answering my own questions, here are my findings so far and a
tentative conclusion.

With or without the patch in this thread, parse_pathspec() behaves
the same way for either CWD or FULL if you feed a non-empty
pathspec with at least one positive element.  IOW, if a caller feeds
a non-empty pathspec and there is no "negative" element involved, it
does not matter if we feed CWD or FULL.

There are only a handful of callers that pass neither preference
bits to parse_pathspec().  Here are my observations on them that
tells me that most of them are OK if we change them to prefer
either CWD or FULL:

 - archive.c::path_exists() feeds a pathspec with a single element
   to see if read_tree_recursive() finds any matching paths, to
   allow its caller to iterate over the original pathspec and see
   if there is a typo (i.e. an element that matches nothing).  It
   should prefer FULL to match what parse_pathspec_arg(), its
   caller, uses.

   The caller probably should refrain from passing ones with
   negative magic.  I.e. "git archive -- t :\!t/perf" errors out
   because checking each element independently in the loop means
   that ":\!t/perf" is checked alone, triggering "there is nothing
   to exclude from".

 - blame.c::find_origin() feeds a pathspec with a single element,
   which is a path in the history and does so as a literal, hence
   no room for "negative" to kick in.

 - builtin/check-ignore.c::check_ignore(), when argc==0, does not
   call parse_pathspec().  It does not take any magic other than
   FROMTOP, so "negative" won't come into the picture.

 - builtin/checkout.c::cmd_checkout(), when argc==0, does not call
   parse_pathspec().  This codepath will get affected by Linus's
   change ("cd t && git checkout :\!perf" would try to check out
   everything except t/perf, but what is a reasonable definition of
   "everything" in the context of this command).  We need to
   decide, and I am leaning towards preferring CWD for this case.

 - revision.c::setup_revisions() calls parse_pathspec() only when
   the caller gave a non-empty pathspec.  This pathspec is used for
   pruning log traversal (e.g. "only show commits that touch these
   paths") and is affected by Linus's change.  It should favor
   FULL.

 - tree-diff.c::try_to_follow_renames() feeds a pathspec with a
   single element as a literal, hence no room for "negative" to
   kick in.

So, I am tempted to suggest us doing the following:

 * Leave a NEEDSWORK comment to archive.c::path_exists() that is
   used for checking the validation of pathspec elements.  To fix it
   properly, we need to be able to skip a negative pathspec to be
   passed to this function by the caller, and to do so, we need to
   expose a helper from the pathspec API that gets a single string
   and returns what magic it has, but that is of lower priority.

 * Retire the PATHSPEC_PREFER_CWD bit and replace its use with the
   lack of the PATHSPEC_PREFER_FULL bit.

 * Keep most of the above callsites that currently do not pass
   CWD/FULL as they are, except the ones that should take FULL (see
   above).

Comments?

Re: Fwd: Possibly nicer pathspec syntax?

[Duy Nguyen]

On Thu, Feb 9, 2017 at 12:39 AM, Junio C Hamano <gitster@pobox.com> wrote:
> Duy Nguyen <pclouds@gmail.com> writes:
>
>> On Wed, Feb 8, 2017 at 12:12 PM, Linus Torvalds
>> <torvalds@linux-foundation.org> wrote:
>>> Two-patch series to follow.
>>
>> glossary-content.txt update for both patches would be nice.
>
> I am no longer worried about it as I saw somebody actually sent
> follow-up patches on this, but I want to pick your brain on one
> thing that is related to this codepath.
>
> We have PATHSPEC_PREFER_CWD and PATHSPEC_PREFER_FULL bits in flags,
> added at fc12261fea ("parse_pathspec: add PATHSPEC_PREFER_{CWD,FULL}
> flags", 2013-07-14), and I think the intent is some commands when
> given no pathspec work on all paths in the current subdirectory
> while others work on the full tree, regardless of where you are.
> "grep" is in the former camp, "log" is in the latter.  And there is
> a check to catch a bug in a caller that sets both.
>
> I am wondering about this hunk (this is from the original commit
> that added it):
>
>         if (!entry) {
>                 static const char *raw[2];
>
> +               if (flags & PATHSPEC_PREFER_FULL)
> +                       return;
> +
> +               if (!(flags & PATHSPEC_PREFER_CWD))
> +                       die("BUG: PATHSPEC_PREFER_CWD requires arguments");
> +
>                 pathspec->items = item = xmalloc(sizeof(*item));
>                 memset(item, 0, sizeof(*item));
>                 item->match = prefix;
>                 ... returns a single entry pathspec to cover cwd ...
>
> The BUG message is given when
>
>  - The command got no pathspec from the caller; and
>  - PATHSPEC_PREFER_FULL is not set; and
>  - PATHSPEC_PREFER_CWD is NOT set.
>
> but the message says that the caller must have args when it sets
> prefer-cwd.  Is this a simple typo?  If so what should it say?
>
>         die("BUG: one of PATHSPEC_PREFER_FULL or _CWD must be set");

Without reading through your next mail, I'd say "BUG: this command
requires arguments".

> Does this third possibility (i.e. a caller is allowed to pass
> "flags" that does not prefer either) exist to support a command
> where the caller MUST have at least one pathspec?  If that were the
> case, this wouldn't be a BUG but an end-user error, e.g.
>
>         die("at least one pathspec element is required");

Or this. Yes. I might have just been defensive at then and kept the
third option open.

> If you know offhand which callers pass neither of the two
> PATHSPEC_PREFER_* bits and remember for what purpose you allowed
> them to do so, please remind me.  I'll keep digging in the meantime.

I don't usually remember what I ate yesterday and this commit was from
2013 :D But I'll see if your findings spark anything in my brain.
-- 
Duy

Re: Fwd: Possibly nicer pathspec syntax?

[Duy Nguyen]

On Thu, Feb 9, 2017 at 4:11 AM, Junio C Hamano <gitster@pobox.com> wrote:
> With or without the patch in this thread, parse_pathspec() behaves
> the same way for either CWD or FULL if you feed a non-empty
> pathspec with at least one positive element.  IOW, if a caller feeds
> a non-empty pathspec and there is no "negative" element involved, it
> does not matter if we feed CWD or FULL.

Yes. Now that you put it that way, it may make more sense to rename
the options to PATHSPEC_DEFAULT_* instead of PATHSPEC_PREFER_*.

>  - builtin/checkout.c::cmd_checkout(), when argc==0, does not call
>    parse_pathspec().  This codepath will get affected by Linus's
>    change ("cd t && git checkout :\!perf" would try to check out
>    everything except t/perf, but what is a reasonable definition of
>    "everything" in the context of this command).  We need to
>    decide, and I am leaning towards preferring CWD for this case.

So far I have seen arguments with single negative pathspec as
examples. What about "cd t; git checkout :^perf :^../Documentation"?
CWD is probably not the best base to be excluded from. Maybe the
common prefix of all negative pathspecs? But things start to get a bit
unintuitive on that road. Or do will still bail out on multiple
negative pathspecs with no positive one?

Also, even with single negative pathspec, what about "cd t; git
checkout :^../ewah"?

> So, I am tempted to suggest us doing the following:
>
>  * Leave a NEEDSWORK comment to archive.c::path_exists() that is
>    used for checking the validation of pathspec elements.  To fix it
>    properly, we need to be able to skip a negative pathspec to be
>    passed to this function by the caller, and to do so, we need to
>    expose a helper from the pathspec API that gets a single string
>    and returns what magic it has, but that is of lower priority.

Side note. There's something else I'm not happy with pathspec handling
in "git archive". Try "cd t; git archive HEAD -- ../Documentation" and
you'll get a very unfriendly error message. But well, no time for it.

>  * Retire the PATHSPEC_PREFER_CWD bit and replace its use with the
>    lack of the PATHSPEC_PREFER_FULL bit.
>
>  * Keep most of the above callsites that currently do not pass
>    CWD/FULL as they are, except the ones that should take FULL (see
>    above).
>
> Comments?

This comes from my experience reading files-backend.c. I didn't
realize '0' in files_downcast(ref_store, 0, "pack_refs"); meant
'submodule not allowed' because apparently I was too lazy to read
prototype. But if was files_downcast(ref_store, NO_SUBMODULE,
"pack_refs"), it would save people's time.

With that in mind, should we keep _CWD the name, so people can quickly
see and guess that it prefers _CWD than _FULL? _CWD can be defined as
zero. It there's mostly as a friendly reminder.

Other than that, yes if killing one flag helps maintenance, I'm for it.

PS. You may notice I favored ^ over ! already. ! was a pain point for
me too but I was too lazy to bring it up and fight for it. Thanks
Linus.
-- 
Duy