patch, 8 partsdiff-delta.c: "diff.h" is not a required include

From: Brandon Casey <drafnel@gmail.com>

Junio,

Here is a re-roll of the work on Solaris which integrates the ideas from you and Jeff. This should replace bc/solaris in pu.

This should allow compiling on Solaris with or without a c99 compiler, GCC or SUNWspro.

Solaris 7 should be able to compile when using GCC and bash.

-brandon

Brandon Casey (7):
  Makefile: use /usr/ucb/install on SunOS platforms rather than
    ginstall
  Makefile: add NEEDS_RESOLV to optionally add -lresolv to compile
    arguments
  diff-delta.c: "diff.h" is not a required include
  On Solaris choose the OLD_ICONV iconv() declaration based on the UNIX
    spec
  git-compat-util.h: tweak the way _XOPEN_SOURCE is set on Solaris
  Makefile: define __sun__ on SunOS
  Makefile: add section for SunOS 5.7

Junio C Hamano (1):
  Makefile: introduce SANE_TOOL_PATH for prepending required elements
    to PATH

 Makefile          |   40 +++++++++++++++++++++++++++++++++++-----
 diff-delta.c      |    1 -
 git-compat-util.h |   17 ++++++++++++++---
 git-sh-setup.sh   |    2 ++
 utf8.c            |    2 +-
 5 files changed, 52 insertions(+), 10 deletions(-)

From: Brandon Casey <drafnel@gmail.com>

We can avoid a GNU dependency by using /usr/ucb/install.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 Makefile |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

From: Brandon Casey <drafnel@gmail.com>

This library is required on Solaris when compiling with NO_IPV6 since hstrerror resides in libresolv. Additionally, Solaris 7 will need it, since inet_ntop and inet_pton reside there too.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 Makefile |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)

From: Brandon Casey <drafnel@gmail.com>

This file (diff.h) provides declarations for some functions that are implemented in diff-delta.c. The SUNWspro C99 compiler complains about it. There is nothing defined in "diff.h" that is required by diff-delta.c, so don't #include it.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 diff-delta.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)

From: Brandon Casey <drafnel@gmail.com>

OLD_ICONV is only necessary on Solaris until UNIX03. This is indicated by the private macro _XPG6 which is set in /usr/include/sys/feature_tests.h.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 Makefile |    3 ---
 utf8.c   |    2 +-
 2 files changed, 1 insertions(+), 4 deletions(-)

From: Brandon Casey <drafnel@gmail.com>

On Solaris, when _XOPEN_EXTENDED is set, its header file forces the programs to be XPG4v2, defeating any _XOPEN_SOURCE setting to say we are XPG5 or XPG6. Also on Solaris, XPG6 programs must be compiled with a c99 compiler, while non XPG6 programs must be compiled with a pre-c99 compiler.

So when compiling on Solaris, always refrain from setting _XOPEN_EXTENDED, and then set _XOPEN_SOURCE to 600 or 500 based on whether a c99 compiler is being used or not.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 git-compat-util.h |   17 ++++++++++++++---
 1 files changed, 14 insertions(+), 3 deletions(-)

From: Brandon Casey <drafnel@gmail.com>

The SUNWspro compiler does not define __sun__ (like GCC does). A check of this macro was recently added to detect compilation on SunOS and to modify the handling of the NO_ICONV and _XOPEN_SOURCE feature macros.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 Makefile |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

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

Some platforms (like SunOS and family) have kept their common binaries at some historical moment in time, and introduced new binaries with modern features in a special location like /usr/xpg4/bin or /usr/ucb. Some of the features provided by these modern binaries are expected and required by git. If the featureful binaries are not in the users path, then git could end up using the less featureful binary and fail.

So provide a mechanism to prepend elements to the users PATH at runtime so the modern binaries will be found.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 Makefile        |   14 ++++++++++++++
 git-sh-setup.sh |    2 ++
 2 files changed, 16 insertions(+), 0 deletions(-)

From: Brandon Casey <drafnel@gmail.com>

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 Makefile |   10 ++++++++++
 1 files changed, 10 insertions(+), 0 deletions(-)

On Fri, Jun 05, 2009 at 06:36:15PM -0500, Brandon Casey wrote:

So this bit me already, and it's only been in next for a day. :) I _already_ have /usr/xpg4/bin in my PATH before /usr/bin, but with this patch, I get it stuck at the _beginning_ of my PATH automagically. Which overrides, against my wishes, the "even more sane than /usr/xpg4/bin" part of my PATH that comes at the beginning.

Specifically, I have "~peff/local/bin" at the beginning of my PATH which contains a 'vi' that points to vim. Running "git rebase -i" now puts /usr/xpg4/bin at the beginning of the PATH (before ~peff/local/bin), which means I end up running the crappy system vi instead. For bonus fun, "git commit" still runs the correct 'vi' because it doesn't happen to be implemented as a shell script.

Am I crazy for not having EDITOR=vim instead of EDITOR=vi? Perhaps. But I wanted to point out that tweaking the PATH behind the user's back does cause surprises in the real world.

-Peff

Jeff King wrote:

Good points. I'm fine with dropping this patch, especially when it causes problems for a real Solaris user, which I'm not.

I don't like that git has a dependency on the user's PATH being set correctly though. That's why I liked the patch. I guess I could modify all the uses of sed and friends to look like $SED and then set SED to /usr/xpg4/bin/sed on Solaris. It doesn't sound like that is necessary in practice though.

btw, this patch does help the test suite when the test suite is run using make. The patch added SANE_TOOL_PATH to PATH in the Makefile.

-brandon

On Mon, Jun 08, 2009 at 08:39:50AM -0500, Brandon Casey wrote:

Let me point out that I'm also not a real Solaris user. These days all I use it for is test-compiling git. So you can take my report with a grain of salt.

Yeah, I think requiring the user's PATH to be set correctly and tweaking the PATH behind the user's back are both unsatisfactory solutions. Using $SED everywhere solves both problems, but would probably be quite annoying to maintain. So I guess it is a matter of picking our poison.

-Peff

Jeff King wrote:

heh.

agreed.

-brandon

Jeff King <peff@peff.net> writes:

In git-sh-setup, we do "unset CDPATH" ourselves to help and protect clueless people, even though "people should have a sane environment". Even though I suspect that anybody who is using Solaris for anything real would not be using /usr/bin tools themselves (i.e. it should not be necessary for us fixing their PATH), there may be people who do not know. I think helping them with path munging falls into the same category, but at the same time, the remedy looks worse than the disease.

We could further uglify the patch like this.

 Makefile        |    5 +++--
 git-sh-setup.sh |   28 +++++++++++++++++++++++++++-
 2 files changed, 30 insertions(+), 3 deletions(-)

On Mon, Jun 08, 2009 at 09:41:49AM -0700, Junio C Hamano wrote:

Wow. That _is_ ugly, but it actually addresses exactly both my concern and Brandon's. I kind of like it.

-Peff

I never received the referenced email. I'll try to extract from gmane and test.

-brandon

Jeff King wrote:

Brandon Casey wrote:

This patch works for me.

-brandon

From: Brandon Casey <drafnel@gmail.com>

When compiling diff-delta.c with the SUNWspro C99 compiler, it complains

    "diff-delta.c", line 314: identifier redeclared: create_delta

There is nothing in "delta.h" that is required by diff-delta.c, so don't include it.

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---
 diff-delta.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)

On Fri, 5 Jun 2009, Brandon Casey wrote:

NAK.

This is common practice to include the header file declaring function prototypes into the file defining the actual function so to make sure the declaration matches with the definition. Deleting that include is actively ignoring a problem instead of fixing the cause of it.

Nicolas

Nicolas Pitre wrote:

It doesn't seem to like the structure being redeclared with a flex array member and being passed as a const argument.

# cat > test.c <<EOF

struct a_struct;

extern void *test_func(const struct a_struct *f);

struct a_struct {
        int a;
        int b;
        char* c[];
};

void *test_func(const struct a_struct *f)
{
        return 0;
}
EOF

# /opt/SUNWspro/bin/c99 -c test.c 
"test.c", line 13: identifier redeclared: test_func
        current : function(pointer to const struct a_struct {int a, int b, array[-1] of pointer to char c}) returning pointer to void
        previous: function(pointer to const struct a_struct {int a, int b, array[-1] of pointer to char c}) returning pointer to void : "test.c", line 4
c99: acomp failed for test.c

If either the flex array is removed from the structure, or const is removed from test_func argument, test.c will compile. Compiling with -O0 doesn't help.

-brandon

On Fri, 5 Jun 2009, Brandon Casey wrote:

What if you define FLEX_ARRAY to 1, or even 0?

If neither of those work then I'd simply remove the const. Generated code should be exactly the same with gcc. There is no const with sizeof_delta_index() which is already inconsistent.

Kind of weird nevertheless.

Nicolas

On Fri, Jun 5, 2009 at 10:10 PM, Nicolas Pitre<nico@cam.org> wrote:

I tried that with my test.c example and '1' works, but not '0'. I'll try setting FLEX_ARRAY to 1 and running git's test suite on Monday.

Yes.

-brandon

From: Brandon Casey <drafnel@gmail.com>

The Sun c99 compiler as recent as version 5.8 Patch 121016-06 2007/08/01 produces an error when compiling diff-delta.c. This source file #includes the delta.h header file which pre-declares a struct which is later defined to contain a flex array member. The Sun c99 compiler fails to compile diff-delta.c and gives the following error:

  "diff-delta.c", line 314: identifier redeclared: create_delta
          current : function(pointer to const struct delta_index {unsigned long memsize, pointer to const void src_buf, unsigned long src_size, unsigned int hash_mask, array[-1] of pointer to struct index_entry {..} hash}, pointer to const void, unsigned long, pointer to unsigned long, unsigned long) returning pointer to void
          previous: function(pointer to const struct delta_index {unsigned long memsize, pointer to const void src_buf, unsigned long src_size, unsigned int hash_mask, array[-1] of pointer to struct index_entry {..} hash}, pointer to const void, unsigned long, pointer to unsigned long, unsigned long) returning pointer to void : "delta.h", line 44
  c99: acomp failed for diff-delta.c

So, avoid using this c99 feature when compiling with the Sun c compilers version 5.8 and older (the most recent version tested).

Signed-off-by: Brandon Casey <drafnel@gmail.com>
---

This should avoid the flex array problems when using the Sun c99 compiler.

This patch is on top of the new bc/solaris (a7a24ee7).

Since this checks the version of the Sun compiler, it should give Sun the opportunity to fix the compiler in newer releases. If someone has Sun Studio 12? where __SUNPRO_C is set to 0x590, maybe they can test.

-brandon

 git-compat-util.h |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

Brandon Casey <casey@nrlssc.navy.mil> writes:

Could you please add this build configuration variable to configure.ac and config.mak.in, to be able to autodetect this situation?

CC-ed Ralf Wildenhues and David Syzdek (who hopefully can produce autoconf patch to squash with this one).

On Sat, Jun 6, 2009 at 2:29 AM, Jakub Narebski<jnareb@gmail.com> wrote:

I'll take a look at it, but autoconf is not a strong suit of mine. Plus, I doubt I will actually be able to test anything on the platforms I have access to that need to set NEEDS_RESOLV, since any autoconf installation is likely to be very old.

Yes please. :)

FYI: Solaris 7 needs -lresov since inet_ntop and inet_pton reside there. Additionally, since NO_IPV6 is set, hstrerror is called in connect.c and hstrerror also resides in libresolv.

On more modern Solaris, inet_ntop and inet_pton reside somewhere else, and since NO_IPV6 does not need to be set, -lresolv is not needed.

-brandon

Check if -lresolv is needed for hstrerror; set NEEDS_RESOLV accordingly, and substitute in config.mak.in.

Signed-off-by: Ralf Wildenhues <Ralf.Wildenhues@gmx.de>
---

* Jakub Narebski wrote on Sat, Jun 06, 2009 at 09:29:34AM CEST:

Completely untested, but also completely mechanical.  HTH.

Cheers, Ralf

 config.mak.in |    1 +
 configure.ac  |    9 +++++++++
 2 files changed, 10 insertions(+), 0 deletions(-)

Brandon Casey wrote:

Just to be clear, this series is built on top of master.

-brandon

Looked good except for 3/8 which I did not quite understand.

Junio C Hamano wrote:

I get this error when compiling using the SUNWspro c99 compiler when delta.h is included in diff-delta.c:

"diff-delta.c", line 314: identifier redeclared: create_delta
        current : function(pointer to const struct delta_index {unsigned long memsize, pointer to const void src_buf, unsigned long src_size, unsigned int hash_mask, array[-1] of pointer to struct index_entry {..} hash}, pointer to const void, unsigned long, pointer to unsigned long, unsigned long) returning pointer to void
        previous: function(pointer to const struct delta_index {unsigned long memsize, pointer to const void src_buf, unsigned long src_size, unsigned int hash_mask, array[-1] of pointer to struct index_entry {..} hash}, pointer to const void, unsigned long, pointer to unsigned long, unsigned long) returning pointer to void : "delta.h", line 44
c99: acomp failed for diff-delta.c
gmake: *** [diff-delta.o] Error 2

I don't see any difference between those two "current" and "previous" statements.

I thought I knew why the error was occurring, but now I don't think I do. There are other function declarations in delta.h that are implemented in diff-delta.c, and those functions are both declared and implemented _before_ create_delta in delta.h and diff-delta.c respectively.

But, there does not seem to be anything declared in delta.h that is required by diff-delta.c.

Maybe the commit message should be shortened to something more like:

   The SUNWspro C99 compiler complains: "identifier redeclared: create_delta" when
   delta.h is included.  There is nothing in "delta.h" that is required by
   diff-delta.c, so don't #include it.

Hmm, well, I just noticed that in the commit message I said "diff.h" everywhere when I meant to say "delta.h". Maybe that is the confusion.

-brandon

On Fri, Jun 05, 2009 at 06:36:08PM -0500, Brandon Casey wrote:

With the exception of 7/8 (which I already complained about separately), these look reasonable to me, and my setup still passes the same tests (though that is perhaps not saying much, as I was already using gcc).

-Peff