Re: [Boost-build] x86 CPU spinning should use pause

Subject: Re: [Boost-build] x86 CPU spinning should use pause
From: Hite, Christopher (Christopher.Hite_at_[hidden])
Date: 2011-10-05 07:26:44

I tried it out and you're right:

void pause() {
        __asm__ __volatile__("pause;");
}

void rep_nop() {
        __asm__ __volatile__( "rep; nop" : : : "memory" );  // copied from BOOST_SMT_PAUSE yield_k.hpp
}

080483c4 <_Z5pausev>:
 80483c4:       55                      push   %ebp
 80483c5:       89 e5                   mov    %esp,%ebp
 80483c7:       f3 90                   pause
 80483c9:       5d                      pop    %ebp
 80483ca:       c3                      ret

080483cb <_Z7rep_nopv>:
 80483cb:       55                      push   %ebp
 80483cc:       89 e5                   mov    %esp,%ebp
 80483ce:       f3 90                   pause
 80483d0:       5d                      pop    %ebp
 80483d1:       c3                      ret

Still the other points in the Intel paper are valid.  If you fail a CAS you should probably loop and read it.  Otherwise if multiple threads are waiting they'll rip the cache line back and forth.

Also the impl calls nanosleep(), which kind of scares me because I would expecting the spin mutex to spin. If this happens often, we should be using OS mutexes.

The time on the OS sleeps is >> 1us, on my Linux ~3ms.  1us is probably way greater than 33 spins you do before calling sleep.

Ok, I realize spinlock is in::detail.  As long as asio doesn't use it, I'm fine.  I've stopped using it.  I'm doing low latency work and stuff like that scares me a bit.  As long as I don't do any atomic_exchange() of shared_ptr, I should be fine.

Chris

-----Original Message-----
From: boost-build-bounces_at_[hidden] [mailto:boost-build-bounces_at_[hidden]] On Behalf Of Jeremiah Willcock
Sent: Tuesday, October 04, 2011 5:35 PM
To: Boost.Build developer's and user's list
Subject: Re: [Boost-build] x86 CPU spinning should use pause

On Tue, 4 Oct 2011, Hite, Christopher wrote:

>
> Sorry for posting to boost-build.  I mean to do users.
>
> According to this Intel forum they take different times:
> http://software.intel.com/en-us/forums/showthread.php?t=48371
> That means they can't be the same instruction.  I'd trust their whitepaper.
>
> Pause and no-op are two different use cases.  No-ops can allow a 
> debugger to put jumps into the code.  This pause instruction means 
> flush out memory operations.

Pause is not the same as the normal nop, but it is the same as rep-nop (they produce the exact same machine code).  See the page that I linked to about that.

-- Jeremiah Willcock

>
>
>
> -----Original Message-----
> From: boost-build-bounces_at_[hidden] 
> [mailto:boost-build-bounces_at_[hidden]] On Behalf Of Jeremiah 
> Willcock
> Sent: Wednesday, September 28, 2011 7:12 PM
> To: Boost.Build developer's and user's list
> Subject: Re: [Boost-build] x86 CPU spinning should use pause
>
> On Wed, 28 Sep 2011, Hite, Christopher wrote:
>
>> #define BOOST_SMT_PAUSE __asm__ __volatile__( "rep; nop" : : : 
>> "memory" );
>>  
>> We should probably use:
>>         __asm__ __volatile__("pause;")
>>  
>> Why? Because one thread polling( or worse CASing) memory, causes multiple overlapping memory operations to be made:
>
> See
> http://stackoverflow.com/questions/7086220/what-does-rep-nop-mean-in-x
> 86-assembly
> -- basically, "rep; nop" is exactly the same instruction as "pause".
>
> -- Jeremiah Willcock
> _______________________________________________
> Unsubscribe & other changes: 
> http://lists.boost.org/mailman/listinfo.cgi/boost-build
>