Subj : Re: CMPXCHG timing
To   : comp.programming.threads
From : Michael Pryhodko
Date : Fri Apr 01 2005 04:04 pm

> > processor reads value from memory, compares it with EAX and if they
are
> > different processor WRITES BACK original value!! Shit!! Who and why
> > decided to do that??? Until someone explains me reasons behind this
> > decision I will consider it as very "bright" and "brilliant" move
from
> > Intel.
>
>     This is well-documented behavior of the CMPXCHG function. The
write is
> unconditional. You should not be complaining about this because if
you want
> atomic behavior, you have to use the LOCK prefix anyway, in which
case the
> write is harmless. If you don't want atomic behavior, implement the
function
> yourself without the extra write.

I DO NOT WANT ATOMIC BEHAVIOR! Please, read carefully more formal
description of algorithm I posted. The whole idea of it was "we do not
need LOCK to do lock".
Main "feature" of algorithm was that ONLY i-th processor could ever
write unique value Pi into lock variable, but with this stupid
implementation of CMPXCHG any processor that tries to enter locked lock
writes Pi of lock owner every CMPXCHG call. This spoils whole idea.
I do not understand why Intel decided to implement CMPXCHG this way,
but from all I learned so far I do not know any good "excuse" for that.


>     By the way, looking at all of your posts on this issue, I think
you're
> totally barking up the wrong tree.

I suggest you to take care next time -- this statement is quite an
insult in my culture.


> Single instructions without the LOCK
> prefix are not atomic with respect to other processors anyway -- if
they
> were, what would the LOCK prefix be for? And you cannot rely on
instruction
> timings for synchronization because of things like cache misses,
speculative
> reads, and so on.

Please read carefully IA manual, especially section where it states
necessary conditions for simple operations to be atomic. For example
store of 32bit value  to 4byte aligned memory address is atomic without
any LOCK (which is forbidden for 'mov').

And by the way, in this example I took 'cache misses' and 'speculative
reads' into account. Read comments carefully and pay attention to
sfences/mfences in the code.


>     I suggest you think carefully about what you're trying to do.

Read carefully before answering. ;)


Bye.
Sincerely yours, Michael.

.