[RFC/PATCH] git-compat-util.h: Don't define NORETURN under __clang__

clang version 1.0 on Debian testing x86_64 defines __GNUC__, but barfs
on `void __attribute__((__noreturn__))'. E.g.:

    usage.c:56:1: error: function declared 'noreturn' should not return [-Winvalid-noreturn]
    }
    ^
    1 diagnostic generated.
    make: *** [usage.o] Error 1

There it's dying on `void __attribute__((__noreturn__)) usagef(const
char *err, ...)' in usage.c, which doesn't return.

Change the header to define NORETURN to nothing under clang. This was
the default behavior for non-GNU and non-MSC compilers already. Having
NORETURN_PTR defined to the GNU C value has no effect on clang
however.

Signed-off-by: Ævar Arnfjörð Bjarmason <redacted>
---

I have no experience with Clang so this may not be sane, but on my
system clang compiles with it and passes all tests. It still spews a
lot of warnings though:
    
    GITGUI_VERSION = 0.12.0.64.g89d61
        * new build flags or prefix
    config.c:297:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    connect.c:151:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    date.c:678:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    diff.c:429:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.

These are valid, as control may not return the advertised type if we
call die() inside them.

    log-tree.c:297:21: warning: field width should have type 'int', but argument has type 'unsigned int' [-Wformat]
                             "Subject: [%s %0*d/%d] ",
                                             ^
    1 diagnostic generated.
    notes.c:632:25: warning: field precision should have type 'int', but argument has type 'unsigned int' [-Wformat]
            strbuf_addf(buf, "%o %.*s%c", mode, path_len, path, '\0');
                                   ^            ~~~~~~~~
    1 diagnostic generated.

Should these (and some below) just cast to (int) ?

    object.c:44:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    parse-options.c:155:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    read-cache.c:1361:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    remote.c:658:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    revision.c:253:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    setup.c:79:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    sideband.c:97:25: warning: field precision should have type 'int', but argument has type 'unsigned int' [-Wformat]
                                            fprintf(stderr, "%.*s", brk + sf, b);
                                                               ^    ~~~~~~~~
    1 diagnostic generated.
    transport.c:1133:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    imap-send.c:548:27: warning: more data arguments than '%' conversions [-Wformat-extra-args]
                               cmd->tag, cmd->cmd, cmd->cb.dlen);
                                                   ^
    1 diagnostic generated.
    Writing perl.mak for Git
    GIT_VERSION = 1.7.2.1.7.gb44c1
    builtin/blame.c:1984:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/bundle.c:67:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/commit.c:832:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/fetch-pack.c:209:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/grep.c:704:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/help.c:58:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/pack-redundant.c:584:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/push.c:252:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    builtin/reflog.c:782:1: warning: control may reach end of non-void function [-Wreturn-type]
    }
    ^
    1 diagnostic generated.
    
 git-compat-util.h |    5 ++++-
 1 files changed, 4 insertions(+), 1 deletions(-)

diff --git a/git-compat-util.h b/git-compat-util.h
index 02a73ee..c651cb7 100644
--- a/git-compat-util.h
+++ b/git-compat-util.h

@@ -183,7 +183,10 @@ extern char *gitbasename(char *);
 #define is_dir_sep(c) ((c) == '/')
 #endif
 
-#ifdef __GNUC__
+#ifdef __clang__
+#define NORETURN
+#define NORETURN_PTR __attribute__((__noreturn__))
+#elif __GNUC__
 #define NORETURN __attribute__((__noreturn__))
 #define NORETURN_PTR __attribute__((__noreturn__))
 #elif defined(_MSC_VER)

-- 
1.7.1

Ævar Arnfjörð Bjarmason venit, vidit, dixit 03.08.2010 15:08:

__GNUC__ should be true if defined, but maybe you still want

+#elif defined(__GNUC__)

instead to make this really equivalent in the GNUC case.

 #define NORETURN __attribute__((__noreturn__))
 #define NORETURN_PTR __attribute__((__noreturn__))
 #elif defined(_MSC_VER)

Michael

On 03.08.2010, at 15:08, Ævar Arnfjörð Bjarmason wrote:

clang version 1.0 on Debian testing x86_64 defines __GNUC__, but barfs
on `void __attribute__((__noreturn__))'. E.g.:

This is a bug in clang 1.0 that was fixed in the llvm 2.7/clang 1.5 timeframe, it didn't
recognize attributes on function pointers properly. clang tries to be as compatible with GCC as
possible so it obviously has to define __GNUC__ ;)

   usage.c:56:1: error: function declared 'noreturn' should not return [-Winvalid-noreturn]
   }
   ^
   1 diagnostic generated.
   make: *** [usage.o] Error 1

It's a "warning which defaults to an error" you can pacify it by passing -Winvalid-noreturn or
-Wno-invalid-noreturn.

I oppose adding workarounds for obsolete versions of clang, Debian should upgrade their packages to
a more recent release. clang 1.0 had all kinds of weird bugs and shouldn't be used anymore.