Re: pthread.h INITIALIZER macros on solaris (solaris9 & 10 need fixes)

[Bruce Korb <bkorb at gnu dot org>]

Kaveh R. Ghazi wrote:
> I know this has come up a few times before :-)

Yes.  I do remember.  :-(  --------- ;-)

> If I can get this working, I'd like to install the testcase to ensure
> it stays working, or at least we notice when it breaks with later
> releases of solaris.  (Crossing my fingers...)  They so helpfully
> change this area for every release. :-(

Even in patches, too, so there is no os version to header version mapping.
You have to look at the version of the header that defines upad64_t
in order to figure out how to patch the pthread.h header.

> The current solaris_mutex_init_2 does a partial job to fix this,
> however it omits fixing the RWLOCK initializer.  Also it doesn't
> trigger correctly on solaris9 because sun rearranged the macro
> definition so the fixinc pattern doesn't match.  I'm also finding that
> solaris10 headers fix the upad64_t by never using the union if the
> compiler is GCC!  So we don't need the upad64_t fix there, but there's
> still a problem with PTHREAD_ONCE_INIT separate from the upad64_t
> issue that is common to all solaris versions I have.  (I have a simple
> fix for that.)

I guess they _tried_ to be helpful.  "almost".  I had access to several
versions at my last company, but only one today.

Anyway, Solaris 8, from sys/types.h:
> #pragma ident   "@(#)types.h    1.66    00/02/14 SMI"
[...]
> #if __STDC__ - 0 == 0 && !defined(_NO_LONGLONG)
> typedef int64_t         pad64_t;
> typedef uint64_t        upad64_t;
> #else
> typedef union {
>         double   _d;
>         int32_t  _l[2];
> } pad64_t;
> typedef union {
>         double   _d;
>         uint32_t _l[2];
> } upad64_t;
> #endif
> $ uname -a
> SunOS vega 5.8 Generic_108528-09 sun4u sparc SUNW,Ultra-4

typedef struct  _pthread_mutex {                /* = mutex_t in synch.h */
        struct {
                uint16_t        __pthread_mutex_flag1;
                uint8_t         __pthread_mutex_flag2;
                uint8_t         __pthread_mutex_ceiling;
                uint32_t        __pthread_mutex_type;
        } __pthread_mutex_flags;
        union {
                struct {
                        uint8_t __pthread_mutex_pad[8];
                } __pthread_mutex_lock64;
                upad64_t __pthread_mutex_owner64;
        } __pthread_mutex_lock;
        upad64_t __pthread_mutex_data;
} pthread_mutex_t;

#define PTHREAD_MUTEX_INITIALIZER       {{0, 0, 0, 0}, {{{0}}}, 0}

typedef struct  _pthread_cond {         /* = cond_t in synch.h */
        struct {
                uint8_t         __pthread_cond_flag[4];
                uint32_t        __pthread_cond_type;
        } __pthread_cond_flags;
        upad64_t __pthread_cond_data;
} pthread_cond_t;

#define PTHREAD_COND_INITIALIZER        {{{0}, 0}, 0}   /* = DEFAULTCV */

typedef struct _pthread_rwlock {        /* = rwlock_t in synch.h */
        int32_t         __pthread_rwlock_readers;
        uint16_t        __pthread_rwlock_type;
        uint16_t        __pthread_rwlock_magic;
        upad64_t        __pthread_rwlock_pad1[3];
        upad64_t        __pthread_rwlock_pad2[2];
        upad64_t        __pthread_rwlock_pad3[2];
} pthread_rwlock_t;

#define PTHREAD_RWLOCK_INITIALIZER      {0, 0, 0, {0, 0, 0}, {0, 0}, {0, 0}}

> I'm playing with fixincludes but I only have access to solaris7, 9 and
> 10 and I don't know if there are multiple versions of pthread.h and
> other headers within an individual solaris release that affect this
> problem.

All the relevant structures are in sys/types.h and the problematic
initializers are in pthread.h.

> If you could let me see how your various solaris boxes define the
> pthread.h macros, and how upad64_t is defined in sys/types.h on the
> same boxes, I'm willing to tackle this issue again for solaris 9 and
> 10. :-)

Here's a really simple approach: remove all initialization values after
the first element.  They are all zero and omitted elements are zeroed.  :)
(Or, even plain ``{}'' as an initializer???)

Cheers - Bruce