[AMDGPU][PATCH 2/3] barriers/memory-fences related additions

[Damien Hilloulin]

Hello everyone,

Sorry for replying late and with this address instead of the other, the 
mail server of my school is down since yesterday so I could not see your 
answers. I have read them with the archives.

I will try to answer to each point:

- No problem, I will move the intrinsics to the other file

- I thought that I could provide a (wrong) first implementation of the 
intrinsics if stating that in the code and in the commits messages. I 
thought that because I think that the current implementation of 
barrier.global() is not good on Evergreen (no WAIT_ACK and all the ACK 
insertion in read/writes). Must I provide a working implementation directly?

Sincerely,

Damien.