Skip site navigation (1) Skip section navigation (2)

Re: Alpha tas() patch

From: Brent Verner <brent(at)rcfile(dot)org>
To: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
Cc: pgsql-hackers <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: Alpha tas() patch
Date: 2000-12-28 14:35:59
Message-ID: 20001228093559.A13865@rcfile.org (view raw or flat)
Thread:
Lists: pgsql-hackerspgsql-patches
On 27 Dec 2000 at 21:37 (-0500), Tom Lane wrote:
| Brent Verner <brent(at)rcfile(dot)org> writes:
| >   This is a revised patch that I sent earlier to allow building
| > pg-7.1 with gcc as well as DEC's cc. I've had good results with this
| > applied. Could some other Alpha users try this out. Even better, could
| > an Alpha asm guru look over the asm that I'm using (instead of the
| > original asm in the file).
| 
| tas() is not supposed to contain a loop.  It can succeed or fail, but
| it should not wait.
| 
| The code now in s_lock.h does seem rather gratuitously obscure about
| the instructions it uses to load constants.  I'd suggest
| 
| static __inline__ int
| tas(volatile slock_t *lock)
| {
|  register slock_t _res;
| 
| __asm__("     ldq   $0, %0           \n\
|               bne   $0, 2f           \n\
|               ldq_l $0, %0           \n\
|               bne   $0, 2f           \n\
|               mov   1, $0            \n\
|               stq_c $0, %0           \n\
|               beq   $0, 2f           \n\
|               mov   0, %1            \n\
|               mb                     \n\
|               jmp   $31, 3f          \n\
|            2: mov   1, %1            \n\
|            3: nop      ": "=m"(*lock), "=r"(_res): :"0");
| 
|     return (int) _res;
| }

another loop-free version of TAS that /seems/ to work as desired. WRT to your
seeing the shutdown lock failure, what are you doing to provoke this bug?, 
because I have not seen it with either version of the TAS that I've used.

	brent


#define TAS(lock) tas_dbv(lock)
#define S_UNLOCK(lock) do { __asm__("mb"); *(lock) = 0; } while (0)

static
__inline__
int
tas(volatile slock_t* __lock)
{
   register slock_t  __rv;
   register slock_t  __temp;

   __asm__
   __volatile__
   (
   "1:  ldq_l %0, %1       \n" /* load (and lock) __lock                 */
   "    and   %0, 1, %2    \n" /* if bit 1 is set, store 1 in __rv       */
   "    bne   %2, 2f       \n" /* if __rv != 0, already locked. leave    */
   "    xor   %0, 1, %0    \n" /* set bit 1 in slock_t                   */
   "    stq_c %0, %1       \n" /* store __lock, only stores if we locked */
   "    mb                 \n" /* memory block (necessary?)              */
   "  2:  nop              \n" /* exit point                             */
     : "=&r" (__temp),
       "=m"  (*__lock),
       "=&r" (__rv)
     : "m"   (*__lock)
   );

   return (int) __rv;
}


In response to

Responses

pgsql-hackers by date

Next:From: Adriaan JoubertDate: 2000-12-28 15:03:26
Subject: Re: Re: Alpha tas() patch
Previous:From: Vadim MikheevDate: 2000-12-28 13:51:10
Subject: New WAL version (CRC & data pages backup)

pgsql-patches by date

Next:From: Adriaan JoubertDate: 2000-12-28 15:03:26
Subject: Re: Re: Alpha tas() patch
Previous:From: Brent VernerDate: 2000-12-28 05:05:42
Subject: Re: [PATCHES] Re: Re: Tuple-valued datums on Alpha (was Re: 7.1 on DEC/Alpha)

Privacy Policy | About PostgreSQL
Copyright © 1996-2014 The PostgreSQL Global Development Group