Re: libstdc++/3584: arm-specific atomic operations not atomic

[Robin Farine <acnrf@dial.eunet.ch>]

The following reply was made to PR libstdc++/3584; it has been noted by GNATS.

From: Robin Farine <acnrf@dial.eunet.ch>
To: Richard.Earnshaw@arm.com
Cc: robin.farine@terminus.org, gcc-gnats@gcc.gnu.org
Subject: Re: libstdc++/3584: arm-specific atomic operations not atomic
Date: 12 Jul 2001 16:08:17 +0200

 Robin Farine <rfarine@halftrack.hq.acn-group.ch> writes:
 
 [...]
 
 > 
 >         @ r0 = address (struct {int lock, int value}), r1 = increment.
 >         ...                     @ prolog
 >         mov     r2, 1
 > 0:
 >         swp     r3, r2, [r0]
 >         teq     r3, #0
 >         beq     1f              @ got the lock
 >         mov     lr, pc
 >         ldr     pc, .thread_yield
 >         b       0b
 > .thread_yield:  .word thread_yield
 > 
 > 1:
 >         @ We now have the lock
 >         ldr     r2, [r0, #4]
 >         add     r2, r2, r1
 >         str     r2, [r0, #4]
 >         str     r3, [r0]        @ Release the lock.
 
 Besides the fact that the label "0:" should *precede* the "mov r2, 1"
 instruction, this version too would allow starvation of a low-priority thread. I
 suppose that a multi-threads and multi-processors safe implementation of
 atomic_add() would look more like this:
 
 void
 atomic_add(volatile _Atomic_word *__mem, int __val)
 {
   int __lock;
   int __tmp = 1;
 
   thread_suspend_scheduler();
 
   __asm__ __volatile__ (
 	"\n"
 
   /* spin while another CPU holds the lock */
 
 	"0:\t"
 	"swp     %0, %1, [%2] \n\t"
 	"teq     %0, #0 \n\t"
 	"bne     0b \n\t"
 
   /* we now have the lock */
 
 	"ldr     %1, [%2, #4] \n\t"
 	"add     %1, %1, %3 \n\t"
 	"str     %1, [%2, #4] \n\t"
 
   /* release the lock */
 
 	"str     %0, [%2] \n\t"
 	""
 	: "=&r"(__lock), "=&r"(__tmp)
 	: "r" (__mem), "r"(__val) 
 	: "cc", "memory");
 
   thread_resume_scheduler();
 }
 
 
 This routine relies on the structured version of _Atomic_word and on nestable
 primitives to suspend/resume threads scheduling, which the POSIX threads API
 doesn't provide, if I remember correctly ...