RE: git on HP NonStop

RE: git on HP NonStop

[Joachim Schmitz]

> From: Junio C Hamano [mailto:gitster@pobox.com]
> Sent: Tuesday, August 14, 2012 4:44 PM
> To: Joachim Schmitz
> Subject: Re: git on HP NonStop
> 
> "Joachim Schmitz" <jojo@schmitz-digital.de> writes:
> 
> > Interesting, I never mentioned Tandem did I, But still you recognized
> > HP NonStop as that.
> 
> No, *you* did in your patch "#ifdef".

Ah, I see.

> > Well, I do care about that platform,  but if you don't, there's not
> > much point in me trying to get Tandem specific patches applied, is it?
> 
> As long as the change is isolated (i.e. compilation without "#define
TANDEM"

__TANDEM actually

> for other people will produce byte-for-byte identical result as before),
and
> cleanly made (i.e. the resulting source code is not littered with "#ifdef
> TANDEM" in many places), I do not think there is a reason not to have such
a
> patchset.

It isn't in many places, only 2 places in git-compat-util.h so far:
/usr/local/bin/diff -EBbu ./git-compat-util.h.orig ./git-compat-util.h
--- ./git-compat-util.h.orig    2012-07-30 15:50:38 -0500
+++ ./git-compat-util.h 2012-08-12 11:26:46 -0500
@@ -74,7 +74,8 @@
 # define _XOPEN_SOURCE 500
 # endif
 #elif !defined(__APPLE__) && !defined(__FreeBSD__) && !defined(__USLC__) &&
\
-      !defined(_M_UNIX) && !defined(__sgi) && !defined(__DragonFly__)
+      !defined(_M_UNIX) && !defined(__sgi) && !defined(__DragonFly__) && \
+      !defined(__TANDEM)
 #define _XOPEN_SOURCE 600 /* glibc2 and AIX 5.3L need 500, OpenBSD needs
600 for S_ISLNK() */
 #define _XOPEN_SOURCE_EXTENDED 1 /* AIX 5.3L needs this */
 #endif
 @@ -98,6 +99,11 @@
 #include <stdlib.h>
 #include <stdarg.h>
 #include <string.h>
+#ifdef __TANDEM
+# include <strings.h> /* for strcasecmp() */
+  typedef int intptr_t; /* not "int *" ?!? */
+  typedef unsigned int uintptr_t; /* not "unsigned int *" ?!? */
+#endif
 #include <errno.h>
 #include <limits.h>
 #include <sys/param.h>

Too much? The 2nd part is not necessary NonStop specific, any idea for a
better way?

There there's Makefile:
/usr/local/bin/diff -EBbu ./Makefile.orig ./Makefile
--- ./Makefile.orig     2012-07-30 15:50:38 -0500
+++ ./Makefile  2012-08-14 06:07:16 -0500
@@ -1297,6 +1297,45 @@
        NO_CURL =
        NO_EXPAT =
 endif
+ifeq ($(uname_S),NONSTOP_KERNEL)
+       CC = cc -c99 # needs some C99 features, "inline" is just one of them
+       CFLAGS= -g -O
+       prefix = /usr/local
+
+       # as detected by './configure'
+       #NO_CURL = YesPlease # missdetected, disabled, see below
+       NEEDS_SSL_WITH_CURL = YesPlease # added manually, see above
+       HAVE_LIBCHARSET_H=YesPlease
+       NEEDS_LIBICONV = YesPlease # needs libiconv first, changed further
down
+       NO_SYS_SELECT_H=UnfortunatelyYes
+       NO_D_TYPE_IN_DIRENT = YesPlease
+       NO_HSTRERROR=YesPlease
+       NO_STRCASESTR=YesPlease
+       NO_FNMATCH_CASEFOLD = YesPlease
+       NO_MEMMEM = YesPlease
+       NO_STRLCPY = YesPlease
+       NO_SETENV = YesPlease
+       NO_UNSETENV = YesPlease
+       NO_MKDTEMP = YesPlease
+       NO_MKSTEMPS = YesPlease
+       OLD_ICONV=UnfortunatelyYes # currently libiconv-1.9.1
+       NO_REGEX=YesPlease # Why? ToDo?
+       NO_PTHREADS=UnfortunatelyYes # ToDo? Using PUT, maybe?
+
+       # our's are in ${prefix}/bin
+       PERL_PATH = ${prefix}/bin/perl
+       PYTHON_PATH = ${prefix}/bin/python
+
+       # not detected (nor checked for) by './configure'
+       COMPAT_CFLAGS += -DSA_RESTART=0 # we don't have SA_RESTART on
NonStop
+       COMPAT_CFLAGS += -DHAVE_STRING_H=1 # needed in
compat/fnmatch/fnmatch.c
+       NO_ST_BLOCKS_IN_STRUCT_STAT = YesPlease # we don't have that on
NonStop
+       NO_NSEC = YesPlease # we don't have that on NonStop
+       NO_PREAD = YesPlease # we could use floss_pread though?
+       NO_MMAP = YesPlease # we could use floss_mmap though?
+       # newly implemented further down
+       NO_POLL = YesPlease # we could use floss_poll though?
+endif
 ifneq (,$(findstring MINGW,$(uname_S)))
        pathsep = ;
        NO_PREAD = YesPlease
@@ -1526,6 +1565,11 @@
        LIB_4_CRYPTO = $(OPENSSL_LINK) -lcrypto
 endif
 endif
+ifndef NO_GETTEXT
+ifndef LIBC_CONTAINS_LIBINTL
+       EXTLIBS += -lintl
+endif
+endif
 ifdef NEEDS_LIBICONV
        ifdef ICONVDIR
                BASIC_CFLAGS += -I$(ICONVDIR)/include
@@ -1538,11 +1582,6 @@
 ifdef NEEDS_LIBGEN
        EXTLIBS += -lgen
 Endif
-ifndef NO_GETTEXT
-ifndef LIBC_CONTAINS_LIBINTL
-       EXTLIBS += -lintl
-endif
-endif
 ifdef NEEDS_SOCKET
        EXTLIBS += -lsocket
 endif
@@ -1591,6 +1630,11 @@
        BASIC_CFLAGS += -DNO_GETTEXT
        USE_GETTEXT_SCHEME ?= fallthrough
 endif
+ifdef NO_POLL
+       NO_SYS_POLL_H = YesPlease
+       COMPAT_CFLAGS += -DNO_POLL -Icompat/win32 # so it finds poll.h
+       COMPAT_OBJS += compat/win32/poll.o
+endif
 ifdef NO_STRCASESTR
        COMPAT_CFLAGS += -DNO_STRCASESTR
        COMPAT_OBJS += compat/strcasestr.o

How das that look to you?

> >> Finally, I would prefer to see any message that is addressed to me as
> >> the project maintainer to be Cc'ed to the list.  I feel 80% of my
> >> time writing
> > this
> >> message wasted, as the contents of it would have helped not just you
> >> but 4 other people who want to get their toes in the development
> >> community, but you robbed them the chance of seeing it.
> >
> > Sorry for having wasted your time, but you could just as easily have
> > CCed the list on your reply, couldn't you?
> 
> No, responding to a private message in public without permission goes
against
> generally accepted netiquette (I am OK if you respond to this or to the
previous
> message from me in public).

OK, understood and accepted. Herewith you have my permission ;-)

> > The use of getrlimit(RLIMIT_NOFILE, ...) vs. sysconf(_SC_OPEN_MAX) in
> > sha1_file.c, I believe the later to be more portable, also it is used
> > in other places already (namely compat/win32/poll.c)
> 
> Careful.  Getting rid of the use of getrlimit() from shared code by
introducing
> our own abstraction to be implemented in compat/* for each platform is
fine,
> but that does not mean sysconf() is acceptable by all platforms (your
finding
> use of sysconf() in compat/win32 merely means Win is OK with it).

Well. NonStop is OK with it too. 
And I've brackets with #ifdef _SC_OPEN_MAX, just like in compat/win23:
/usr/local/bin/diff -EBbu ./sha1_file.c.orig ./sha1_file.c
--- ./sha1_file.c.orig  2012-07-30 15:50:38 -0500
+++ ./sha1_file.c       2012-08-14 02:42:03 -0500
@@ -747,6 +747,9 @@
                return error("packfile %s index unavailable", p->pack_name);

        if (!pack_max_fds) {
+#ifdef _SC_OPEN_MAX
+               unsigned int max_fds = sysconf(_SC_OPEN_MAX);
+#else
                struct rlimit lim;
                unsigned int max_fds;

@@ -754,6 +757,7 @@
                        die_errno("cannot get RLIMIT_NOFILE");

                max_fds = lim.rlim_cur;
+#endif

                /* Save 3 for stdin/stdout/stderr, 22 for work */
                if (25 < max_fds)
 
Can it get any better?

> > The use of curl_easy_strerror() in help.c without checking for a
> > proper curl version (>=7.12.0)
> 
> I do not recall the minimum version of curl we support offhand, but if
there is a
> clean workaround that would be good.

Well, I really have none but to disable that part:
/usr/local/bin/diff -EBbu ./http.c.orig ./http.c
--- ./http.c.orig       2012-07-30 15:50:38 -0500
+++ ./http.c    2012-08-14 02:49:43 -0500
@@ -805,10 +805,12 @@
                                ret = HTTP_REAUTH;
                        }
                } else {
+#if LIBCURL_VERSION_NUM >= 0x070c00 /* no curl_easy_strerror() before that
*/
                        if (!curl_errorstr[0])
                                strlcpy(curl_errorstr,
 
curl_easy_strerror(results.curl_result),
                                        sizeof(curl_errorstr));
+#endif
                        ret = HTTP_ERROR;
                }
        } else {


It certainly is a workaround, whether it is clean or clean enough for now is
to you to decide ;-)

Bye, Jojo

RE: git on HP NonStop

[Jan Engelhardt]

On Tuesday 2012-08-14 17:52, Joachim Schmitz wrote:
> @@ -98,6 +99,11 @@
> #include <stdlib.h>
> #include <stdarg.h>
> #include <string.h>
>+#ifdef __TANDEM
>+# include <strings.h> /* for strcasecmp() */
>+  typedef int intptr_t; /* not "int *" ?!? */
>+  typedef unsigned int uintptr_t; /* not "unsigned int *" ?!? */

Of course not. intptr_t is an integral value capable of holding
a pointer; it is not a pointer to int (because that would really
be redundant to int*.)

RE: git on HP NonStop

[Joachim Schmitz]

> From: Jan Engelhardt [mailto:jengelh@inai.de]
> Sent: Sunday, August 19, 2012 6:26 PM
> To: Joachim Schmitz
> Cc: 'Junio C Hamano'; git@vger.kernel.org
> Subject: RE: git on HP NonStop
> 
> 
> On Tuesday 2012-08-14 17:52, Joachim Schmitz wrote:
> > @@ -98,6 +99,11 @@
> > #include <stdlib.h>
> > #include <stdarg.h>
> > #include <string.h>
> >+#ifdef __TANDEM
> >+# include <strings.h> /* for strcasecmp() */
> >+  typedef int intptr_t; /* not "int *" ?!? */
> >+  typedef unsigned int uintptr_t; /* not "unsigned int *" ?!? */
> 
> Of course not. intptr_t is an integral value capable of holding a pointer;
it is not
> a pointer to int (because that would really be redundant to int*.)

OK, thanks for the clarification.

Another issue I stumbled across:
There are numerous places (well, some 10) were something like the following
is done

int var = var;
char *othervar = othervar;

Here this leads to Compiler warnings  'variable "var" is used before its
value is set' on NonStop. This self-initialization seems to be a GCC
extension (?), but even gcc has a -Winit-self option to warn about this.
Shouldn't that better be like the following?

int var = 0;
char *othervar = NULL;

What is the reason for using that self-init stuff? I don't think it is
really portable, is it?

Bye, Jojo

Re: git on HP NonStop

[Johannes Sixt]

Am 8/20/2012 12:36, schrieb Joachim Schmitz:
> int var = var;
> char *othervar = othervar;
> 
> ... 
>
> What is the reason for using that self-init stuff? I don't think it is
> really portable, is it?

It is used to avoid "var may be used uninitialized" warnings for some
compilers. Officially (according to the C standard), it is undefined
behavior. But I've observed considerable resistance by Junio to fix this
properly. Therefore, unless you can show that your compiler generates
unusable code you better live with the "self-initialization" warnings.

-- Hannes

RE: git on HP NonStop

[Joachim Schmitz]

> From: Johannes Sixt [mailto:j.sixt@viscovery.net]
> Sent: Monday, August 20, 2012 12:57 PM
> To: Joachim Schmitz
> Cc: 'Jan Engelhardt'; 'Junio C Hamano'; git@vger.kernel.org
> Subject: Re: git on HP NonStop
> 
> Am 8/20/2012 12:36, schrieb Joachim Schmitz:
> > int var = var;
> > char *othervar = othervar;
> >
> > ...
> >
> > What is the reason for using that self-init stuff? I don't think it is
> > really portable, is it?
> 
> It is used to avoid "var may be used uninitialized" warnings for some
compilers.

Well, it results in a similar warning on NonStop. "var is used before it is
set" and I think this is equally bad.
In either case we don't know what the content of that var is.

E.g. in wt_status.c the variable 'status' is set at only one place, but
later it is switched on. If lucky we get to the default case and die.
So why not just 
int status = 0;

> Officially (according to the C standard), it is undefined behavior. 

Yes, I had that suspicion. Not good to rely in this...

> But I've observed considerable resistance by Junio to fix this properly. 

What's the  reason behind that?

> Therefore, unless
> you can show that your compiler generates unusable code you better live
with
> the "self-initialization" warnings.

So far I can't, so I guess I'll have to live with the warnings, but don't
quite like it.
 
> -- Hannes

Bye, Jojo

Re: git on HP NonStop

[Junio C Hamano]

Johannes Sixt <j.sixt@viscovery.net> writes:

> Am 8/20/2012 12:36, schrieb Joachim Schmitz:
>> int var = var;
>> char *othervar = othervar;
>> 
>> ... 
>>
>> What is the reason for using that self-init stuff? I don't think it is
>> really portable, is it?
>
> It is used to avoid "var may be used uninitialized" warnings for some
> compilers. Officially (according to the C standard), it is undefined
> behavior. But I've observed considerable resistance by Junio to fix this
> properly.

I had resisted

	-	int foo = foo;
        +	int foo = 0;

in the past.  If some compiler is not seeing that "foo" is never
used uninitialized, such a compiler will generate an unnecessary
initialization, so it is not a _proper_ fix anyway (in fact, I do
not think a proper fix exists, short of simplifying the code so that
less sophisticated compilers can see that "foo" is never used
uninitialized).

So, no, I never resisted a "proper" fix.  I resisted swapping an
unsatisfactory workaround with another.

Between the two unsatisfactory workarounds, the latter (explicit
and unnecessary assignment to an innocuous value) is lessor of two
evils, so I do not particularly mind it, though.  Indeed, I think
more recent history shows that we have such changes.

RE: git on HP NonStop

[Joachim Schmitz]

> From: Junio C Hamano [mailto:gitster@pobox.com]
> Sent: Monday, August 20, 2012 6:30 PM
> To: Johannes Sixt
> Cc: Joachim Schmitz; 'Jan Engelhardt'; git@vger.kernel.org
> Subject: Re: git on HP NonStop
> 
> Johannes Sixt <j.sixt@viscovery.net> writes:
> 
> > Am 8/20/2012 12:36, schrieb Joachim Schmitz:
> >> int var = var;
> >> char *othervar = othervar;
> >>
> >> ...
> >>
> >> What is the reason for using that self-init stuff? I don't think it
> >> is really portable, is it?
> >
> > It is used to avoid "var may be used uninitialized" warnings for some
> > compilers. Officially (according to the C standard), it is undefined
> > behavior. But I've observed considerable resistance by Junio to fix
> > this properly.
> 
> I had resisted
> 
> 	-	int foo = foo;
>         +	int foo = 0;
> 
> in the past.  If some compiler is not seeing that "foo" is never used
uninitialized,
> such a compiler will generate an unnecessary initialization, so it is not
a
> _proper_ fix anyway (in fact, I do not think a proper fix exists, short of
> simplifying the code so that less sophisticated compilers can see that
"foo" is
> never used uninitialized).
> 
> So, no, I never resisted a "proper" fix.  I resisted swapping an
unsatisfactory
> workaround with another.
> 
> Between the two unsatisfactory workarounds, the latter (explicit and
> unnecessary assignment to an innocuous value) is lessor of two evils, so I
do not
> particularly mind it, though.  Indeed, I think more recent history shows
that we
> have such changes.

OK, so let's have a look at code, current git,
builtin/cat-file.c, line 196:
        void *contents = contents;

This variable is set later in an if branch (if (print_contents == BATCH),
but not in the else branch. It is later used always under the same condition
as the one under which it is set.
Apparently is is malloc_d storage (there a "free(content);"), so there's no
harm al all in initializing it with NULL, even if it only appeases a stupid
compiler.

The next one, builtin/fast-export.c, line 486:
                struct commit *commit = commit;
it is set in a switch statement, but not in every case, as far as I can see.
Hmm, maybe it is, and I just get lost in the code
And it is used directly after the switch, hopefully set to something
reasonable.
Why take the risk and not set it to NULL?

Next one, builtin/rev-list.c, line 390:
                int reaches = reaches, all = all;

                revs.commits = find_bisection(revs.commits, &reaches, &all,
                                              bisect_find_all);

Seem pretty pointless to initialize them, provided find_bisection doesn't
read them. Does it?
I'm too Next one, lazy to check... I'd just set them to 0 and stop worrying.

Next one, fast-import.c, line 2268:
        struct object_entry *oe = oe;

os gets set in en if and an else branch, but not in then intermediate else
if branch!
It is checked for !NULL later, so it should really get initialized to NULL
in the first place!

Same file, line 2437, same variable name, same story!
Same file, line 2616, variable e, it is used in an if branch but set after
that! 
Same file again, line 2917, variable oe again. Same story as above.

Next file, ll-merge.c, line
        static const struct ll_merge_options default_opts;
Somewhat different story here, compiler warning claims " const variable
"default_opts" requires an initializer"
Possible fix:
        static const struct ll_merge_options default_opts = {0};

next file, match-trees.c, line 75ff:
                const unsigned char *elem1 = elem1;
                const unsigned char *elem2 = elem2;
                const char *path1 = path1;
                const char *path2 = path2;
                unsigned mode1 = mode1;
                unsigned mode2 = mode2;
Some get set, some not, depending on code path, but all get used, with
possibly bogus content.

Next file, merge-recursive.c, line 1903:
        struct tree *mrtree = mrtree;
passed on my address to another function, which hopefully knows how to treat
it. It woult be learer and simpler to just have
        struct tree *mrtree = NULL;
wouldn't it?

Next file, run-command.c, line 272:
        int failed_errno = failed_errno;
Set deeply nested in some cases. Seems to be used reasonably, but again, why
take chanses= 0 is a goot errno ;-)

Next file, submodule.c, line 265:
        struct commit *left = left, *right = right;
As far as I can see it is not set properly before it gets used in some
cases.

Next filen, transport.c, line 106:
                int cmp = cmp, len;
I seem to see code paths whet it is used without being set properly

Next file, vcs-svn/svndiff.c, line 299.... oh, that one has been fixed and
initialized to -1.

Next (and last) file, wt-status.c, line 267:
        int status = status;
It apparently does get set properly before use. So here it once may once
again just make a compiler happy to set it to 0.

Bye, Jojo

Re: git on HP NonStop

[Junio C Hamano]

"Joachim Schmitz" <jojo@schmitz-digital.de> writes:

> OK, so let's have a look at code, current git,
> builtin/cat-file.c, line 196:
>         void *contents = contents;
>
> This variable is set later in an if branch (if (print_contents == BATCH),
> but not in the else branch. It is later used always under the same condition
> as the one under which it is set.
> Apparently is is malloc_d storage (there a "free(content);"), so there's no
> harm al all in initializing it with NULL, even if it only appeases a stupid
> compiler.

It actually is harmful.  See below.

> The next one, builtin/fast-export.c, line 486:
>                 struct commit *commit = commit;
> it is set in a switch statement, but not in every case, as far as I can see.
> Hmm, maybe it is, and I just get lost in the code
> And it is used directly after the switch, hopefully set to something
> reasonable.
> Why take the risk and not set it to NULL?

Ditto.

> Next one, builtin/rev-list.c, line 390:
>                 int reaches = reaches, all = all;
>
>                 revs.commits = find_bisection(revs.commits, &reaches, &all,
>                                               bisect_find_all);
>
> Seem pretty pointless to initialize them, provided find_bisection doesn't
> read them. Does it?

That is why they are not initializations but marks to the compiler
to signal "you may be stupid enough to think they are used before
initialized or assigned, but that is not the case".  Initializing
them would be pointless.

> Next one, fast-import.c, line 2268:
>         struct object_entry *oe = oe;
>
> os gets set in en if and an else branch, but not in then intermediate else
> if branch!

Look again.  If the recent code is too complex for you to
understand, go back to 10e8d68 (Correct compiler warnings in
fast-import., 2007-02-06) and read the function.

The control flow of the early part of that function dictates that
either oe is assigned *or* inline_data is set to 1.  When inline_data
is false, oe is always set.

The compiler was too stupid to read that, and that is why the
(confusing) idiom to mark it for the stupid compiler was used.

There are a few reasons why I do not think this self-assignment
idiom or initializing the variable to an innocuous-looking random
value is a particularly good thing to do when you see compiler
warnings.

If the compiler suspects the variable might be unused, you should
always look at it and follow the flow yourself.  Once you know it is
a false alarm, you can use the idiom to squelch the warning, and it
at the same serves as a note to others that you verified the flow
and made sure it is a false warning.

When the next person wants to touch the code, if the person knows
the use of the idiom, it only serves as a warning to be extra
careful not to introduce a new codepath that reads the variable
without setting, as the compiler no longer helps him.  If the person
who touches the code is as clueless as the compiler and cannot
follow the codepath to see the variable is never used uninitialized,
the result will be a lot worse.

That is the reason why I do not think the idiom to squelch the
compiler is such a good thing.  Careless people touch the code, so
"oe = oe" initialization carefully placed in the original version
does not necessarily stay as a useful documentation.

But if you use "oe = NULL" as a way to squelch the warning in the
first place, it is no better than "oe = oe".  In a sense, it is even
worse, as it just looks like any other initialization and gives a
false impression that the remainder of the code is written in such a
way that it tolerates oe being NULL in any codepath, or there is
some assignment before that before the code reaches places where oe
cannot be NULL.  That is different from what "oe = oe" initializaion
documents---in the codepath protected by "if (inline_data)", it
isn't just "oe can safely be NULL there"; instead it is "oe can
safely be *any* value there, because we don't use it".  Of course,
if you explicitly initialized oe to NULL, even if you introduce a
codepath where oe cannot be NULL later, you won't get a warning from
the compiler, so it is no better than "oe = oe".

And that is the reason why I do not think initialization to an
innocuous-looking random value (e.g. NULL) is a good answer, either.

When both are not good, replacing "oe = oe" with "oe = NULL" didn't
make much sense, especially when the former _could_ be used better
by more careful people.  And that is the resistance J6t remembers.

But when recent compilers started to warn "oe = oe" that itself is
undefined, the equation changed.  The idiom ceased to be a way to
squelch the incorrect compiler warning (which was the primary point
of its use---the documentation value is secondary, as what we
document is "we are squelching a false alarm", but we no longer are
squelching anything).

See 4a5de8d (vcs-svn: avoid self-assignment in dummy initialization
of pre_off, 2012-06-01), and 58ebd98 (vcs-svn/svndiff.c: squelch
false "unused" warning from gcc, 2012-01-27) that it updated, for an
example of how we are migrating away from "oe = oe" idiom.

In any case, I already said initializing to 0 is OK in my previous
message, I think, so if you are seeing warning from the compiler on
uninitialized variables in your new code, you know you should

 (1) first follow the codepath to make sure it is a false alarm (if
     it isn't, fix the code);

 (2) then see if you can simplify your code so that it is crystal
     clear even to the stupid compiler that you are not using an
     uninitialized variable; and

 (3) if you can't, then initialize it to some known bad value,
     avoiding the "oe = oe" idiom.

Is there anything more to discuss on this topic?

RE: git on HP NonStop

[Joachim Schmitz]

> -----Original Message-----
> From: Junio C Hamano [mailto:gitster@pobox.com]
> Sent: Tuesday, August 21, 2012 4:06 AM
> To: Joachim Schmitz
> Cc: 'Johannes Sixt'; 'Jan Engelhardt'; git@vger.kernel.org
> Subject: Re: git on HP NonStop
> 
> "Joachim Schmitz" <jojo@schmitz-digital.de> writes:
> 
> > OK, so let's have a look at code, current git, builtin/cat-file.c,
> > line 196:
> >         void *contents = contents;
> >
> > This variable is set later in an if branch (if (print_contents ==
> > BATCH), but not in the else branch. It is later used always under the
> > same condition as the one under which it is set.
> > Apparently is is malloc_d storage (there a "free(content);"), so
> > there's no harm al all in initializing it with NULL, even if it only
> > appeases a stupid compiler.
> 
> It actually is harmful.  See below.

Harmful to initialize with NULL or to use that undefined behavoir?

I checked what our compiler does here: after having warned about "vlues us
used before it is set: it actually dies seem to have initializes the value
to 0 resp. NULL.
So here there's no harm done in avoiding undefined behavior and set it to 0
resp NULL in the first place.

> > The next one, builtin/fast-export.c, line 486:
> >                 struct commit *commit = commit; it is set in a switch
> > statement, but not in every case, as far as I can see.
> > Hmm, maybe it is, and I just get lost in the code And it is used
> > directly after the switch, hopefully set to something reasonable.
> > Why take the risk and not set it to NULL?
> 
> Ditto.
> 
> > Next one, builtin/rev-list.c, line 390:
> >                 int reaches = reaches, all = all;
> >
> >                 revs.commits = find_bisection(revs.commits, &reaches,
&all,
> >                                               bisect_find_all);
> >
> > Seem pretty pointless to initialize them, provided find_bisection
> > doesn't read them. Does it?
> 
> That is why they are not initializations but marks to the compiler to
signal "you
> may be stupid enough to think they are used before initialized or
assigned, but
> that is not the case".  Initializing them would be pointless.
> 
> > Next one, fast-import.c, line 2268:
> >         struct object_entry *oe = oe;
> >
> > os gets set in en if and an else branch, but not in then intermediate
> > else if branch!
> 
> Look again.  If the recent code is too complex for you to understand, go
back to
> 10e8d68 (Correct compiler warnings in fast-import., 2007-02-06) and read
the
> function.
> 
> The control flow of the early part of that function dictates that either
oe is
> assigned *or* inline_data is set to 1.  When inline_data is false, oe is
always
> set.
> 
> The compiler was too stupid to read that, and that is why the
> (confusing) idiom to mark it for the stupid compiler was used.
> 
> There are a few reasons why I do not think this self-assignment idiom or
> initializing the variable to an innocuous-looking random value is a
particularly
> good thing to do when you see compiler warnings.
> 
> If the compiler suspects the variable might be unused, you should always
look
> at it and follow the flow yourself.  Once you know it is a false alarm,
you can
> use the idiom to squelch the warning, and it at the same serves as a note
to
> others that you verified the flow and made sure it is a false warning.
> 
> When the next person wants to touch the code, if the person knows the use
of
> the idiom, it only serves as a warning to be extra careful not to
introduce a new
> codepath that reads the variable without setting, as the compiler no
longer
> helps him.  If the person who touches the code is as clueless as the
compiler
> and cannot follow the codepath to see the variable is never used
uninitialized,
> the result will be a lot worse.
> 
> That is the reason why I do not think the idiom to squelch the compiler is
such a
> good thing.  Careless people touch the code, so "oe = oe" initialization
carefully
> placed in the original version does not necessarily stay as a useful
> documentation.
> 
> But if you use "oe = NULL" as a way to squelch the warning in the first
place, it
> is no better than "oe = oe".  In a sense, it is even worse, as it just
looks like any
> other initialization and gives a false impression that the remainder of
the code
> is written in such a way that it tolerates oe being NULL in any codepath,
or
> there is some assignment before that before the code reaches places where
oe
> cannot be NULL.  That is different from what "oe = oe" initializaion
documents--
> -in the codepath protected by "if (inline_data)", it isn't just "oe can
safely be
> NULL there"; instead it is "oe can safely be *any* value there, because we
don't
> use it".  Of course, if you explicitly initialized oe to NULL, even if you
introduce
> a codepath where oe cannot be NULL later, you won't get a warning from the
> compiler, so it is no better than "oe = oe".
> 
> And that is the reason why I do not think initialization to an
innocuous-looking
> random value (e.g. NULL) is a good answer, either.
> 
> When both are not good, replacing "oe = oe" with "oe = NULL" didn't make
> much sense, especially when the former _could_ be used better by more
> careful people.  And that is the resistance J6t remembers.
> 
> But when recent compilers started to warn "oe = oe" that itself is
undefined,
> the equation changed.  The idiom ceased to be a way to squelch the
incorrect
> compiler warning (which was the primary point of its use---the
documentation
> value is secondary, as what we document is "we are squelching a false
alarm",
> but we no longer are squelching anything).
> 
> See 4a5de8d (vcs-svn: avoid self-assignment in dummy initialization of
pre_off,
> 2012-06-01), and 58ebd98 (vcs-svn/svndiff.c: squelch false "unused"
warning
> from gcc, 2012-01-27) that it updated, for an example of how we are
migrating
> away from "oe = oe" idiom.
> 
> In any case, I already said initializing to 0 is OK in my previous
message, I think,
> so if you are seeing warning from the compiler on uninitialized variables
in
> your new code, you know you should
> 
>  (1) first follow the codepath to make sure it is a false alarm (if
>      it isn't, fix the code);
> 
>  (2) then see if you can simplify your code so that it is crystal
>      clear even to the stupid compiler that you are not using an
>      uninitialized variable; and
> 
>  (3) if you can't, then initialize it to some known bad value,
>      avoiding the "oe = oe" idiom.
> 
> Is there anything more to discuss on this topic?

Which of the ones I told you about should get fixed?

Bye, Jojo

Re: git on HP NonStop

[Andreas Ericsson]

On 08/22/2012 06:38 PM, Joachim Schmitz wrote:
> 
> 
>> -----Original Message-----
>> From: Junio C Hamano [mailto:gitster@pobox.com]
>> Sent: Tuesday, August 21, 2012 4:06 AM
>> To: Joachim Schmitz
>> Cc: 'Johannes Sixt'; 'Jan Engelhardt'; git@vger.kernel.org
>> Subject: Re: git on HP NonStop
>>
>> "Joachim Schmitz" <jojo@schmitz-digital.de> writes:
>>
>>> OK, so let's have a look at code, current git, builtin/cat-file.c,
>>> line 196:
>>>          void *contents = contents;
>>>
>>> This variable is set later in an if branch (if (print_contents ==
>>> BATCH), but not in the else branch. It is later used always under the
>>> same condition as the one under which it is set.
>>> Apparently is is malloc_d storage (there a "free(content);"), so
>>> there's no harm al all in initializing it with NULL, even if it only
>>> appeases a stupid compiler.
>>
>> It actually is harmful.  See below.
> 
> Harmful to initialize with NULL or to use that undefined behavoir?
> 
> I checked what our compiler does here: after having warned about "vlues us
> used before it is set: it actually dies seem to have initializes the value
> to 0 resp. NULL.
> So here there's no harm done in avoiding undefined behavior and set it to 0
> resp NULL in the first place.
> 

There is harm in tricking future programmers into thinking that the
initialization actually means something, which some of them do.

It's unlikely that you're the one to maintain that code forever, and
the "var = var" idiom is used widely within git with a clear meaning
as a hint to programmers who read a bit of git code. If they aren't
used to that idiom, they usually investigate it in the code and
pretty quickly realize that what it means.


-- 
Andreas Ericsson                   andreas.ericsson@op5.se
OP5 AB                             www.op5.se
Tel: +46 8-230225                  Fax: +46 8-230231

Considering the successes of the wars on alcohol, poverty, drugs and
terror, I think we should give some serious thought to declaring war
on peace.

RE: git on HP NonStop

[Joachim Schmitz]

> From: Andreas Ericsson [mailto:ae@op5.se]
> Sent: Thursday, August 23, 2012 10:24 AM
> To: Joachim Schmitz
> Cc: 'Junio C Hamano'; 'Johannes Sixt'; 'Jan Engelhardt'; git@vger.kernel.org
> Subject: Re: git on HP NonStop
> 
> On 08/22/2012 06:38 PM, Joachim Schmitz wrote:
> >
> >
> >> -----Original Message-----
> >> From: Junio C Hamano [mailto:gitster@pobox.com]
> >> Sent: Tuesday, August 21, 2012 4:06 AM
> >> To: Joachim Schmitz
> >> Cc: 'Johannes Sixt'; 'Jan Engelhardt'; git@vger.kernel.org
> >> Subject: Re: git on HP NonStop
> >>
> >> "Joachim Schmitz" <jojo@schmitz-digital.de> writes:
> >>
> >>> OK, so let's have a look at code, current git, builtin/cat-file.c,
> >>> line 196:
> >>>          void *contents = contents;
> >>>
> >>> This variable is set later in an if branch (if (print_contents ==
> >>> BATCH), but not in the else branch. It is later used always under the
> >>> same condition as the one under which it is set.
> >>> Apparently is is malloc_d storage (there a "free(content);"), so
> >>> there's no harm al all in initializing it with NULL, even if it only
> >>> appeases a stupid compiler.
> >>
> >> It actually is harmful.  See below.
> >
> > Harmful to initialize with NULL or to use that undefined behavoir?
> >
> > I checked what our compiler does here: after having warned about "vlues us
> > used before it is set: it actually dies seem to have initializes the value
> > to 0 resp. NULL.
> > So here there's no harm done in avoiding undefined behavior and set it to 0
> > resp NULL in the first place.
> >
> 
> There is harm in tricking future programmers into thinking that the
> initialization actually means something, which some of them do.

Hmm, OK, I can agree to that.

> It's unlikely that you're the one to maintain that code forever, 

It is unlike for me to ever have to maintain this code.
Currently that's Junio's job and I won't apply for in ;-)

> and the "var = var" idiom is used widely within git 

This is overstating it a bit. I went thru the entire code and reported all places I could find in an earlier email.
I went back and counted: It is used in 11 files, at 15 places, for 21 variables. 
OK, I may have missed  a few more that were in code paths my compiler didn't see, but still some 21+ isn't really much.

>with a clear meaning

Only if you call undefined behavior a  'clear meaning"!

> as a hint to programmers who read a bit of git code. If they aren't
> used to that idiom, they usually investigate it in the code and
> pretty quickly realize that what it means.

Whether I realize what it means, is irrelevant, my compiler does not and warns about it, and as per the ANSI/ICO C standard it
invokes undefined behavior.

If a proper initialization is meaningless for these cases, don't do them at all, let the stupid compiler complain about it and the
clever programmer check whether the warning is useful, but don't avoid a compiler warning on one compiler by introducing undefined
behavior and provoke a compiler warning on another.

Bye, Jojo