[RFC] libstdc++ atomicity, x86 agnostic version

[Paolo Carlini]

Gabriel Dos Reis wrote:

>Why not just ditch that ref from foo()?
>
Yeah, thanks Gaby for you feedback.
I already did that and tried to simulate the overall effect of Benjamin's
work by simply removing the inline keywork from the current
__exchange_and_add and __atomic_add.

These are the numbers (-O2, P4-2400) (Jakub_am is without the &)

current
=======
Jakub (50000000 iters)
5.590u 0.000s 0:05.74 97.3%     0+0k 0+0io 160pf+0w

Jakub_am (50000000)
11.250u 0.000s 0:11.57 97.2%    0+0k 0+0io 160pf+0w

patched
=======
Jakub (50000000)
6.370u 0.000s 0:06.55 97.2%     0+0k 0+0io 162pf+0w

Jakub_am (50000000)
12.600u 0.000s 0:12.96 97.2%    0+0k 0+0io 162pf+0w

So, for this kind of testcase, on the x86 architecture, is a loss of
about 10-12%.

I don't know if we want to live with that... Gaby? Opinions?

Paolo.