Re: notes on current atomics config and usage

[Paolo Carlini <pcarlini at suse dot de>]

Hi Benjamin,

> Here's another approach.
>
> The attached patch clarifies the atomics config. In a nutshell, if
> there exists a config/cpu/*/atomicity.h file, it's used. If there
> isn't, either the builtins are used (and inlined) if they are proved to
> exist, or the mutex fall-back is used. 
>
> There should never be the case where the cpu has both builtins and a
> configure atomicity.h. Thus, this patch removes
> config/cpu/i486 because the builtin atomics for this cpu are better.
>
> As-is, this changes the default behavior for default x86 compiles. What
> happens? Well, say the host is i686, check. But, the i386 atomicity.h
> is skipped over, and there is no i686 atomicity.h, since there are
> actually builtins for this host. However, the check for builtins fails,
> since the default -march for this host is i386. Thus, mutex-ville.
>  
I'm not sure to understand: you mean, by default we are going to use 
mutexes?!? Sorry, but if this what you are really saying, it's 
unacceptable performance-wise: by default everyone (on x86_64 too!) will 
notice a huge regression, please-please let's not do that. In my 
opinion, on i686 by default we should simply provide out of line 
__atomic_add and __exchange_and_add implemented in terms of the builtins 
(we should use builtins instead of assembly, like you are correctly 
suggesting elsewhere).

Note essentially the same issue affects Sparc (Jakub prepared the 
builtins, only they are not exploited by the current configury): 
sufficiently recent machines by default should have out of line code 
using the builtins.

> I did some performance testing, original vs. patch +
> CXXFLAGS="-march=i686". Results are attached. Surprisingly, there is
> some slight degradation,
Well, I'm seeing more-than-slight degradations, some numbers, in 
string_append for example, are horrible (10x !?!), I cannot believe we 
are talking about the same machine!?! Are you really-really sure we are 
talking about the right config? Not even mutexes would explain such huge 
differences...

Benjamin, I would be happy to carry out some performance tests and 
whatelse, but the patch doesn't apply cleanly here, lots of problems... 
Anyway, in a few days I will be away for a week of vacations: please, 
please, let's proceed carefully, separate well tested steps, in case we 
can leave alone the i386/i686 (and Sparc v7/newer) case for now, and 
deal with it separately.

Paolo.