By: Gabriele Svelto (gabriele.svelto.delete@this.gmail.com), October 17, 2006 1:06 am
Room: Moderated Discussions
David Kanter (dkanter@realworldtech.com) on 10/17/06 wrote:
---------------------------
>I suspect that branch slots must terminate a group in the POWER6. They might let
>you use the 5th slot for something else now though...
>
>I just think having basic blocks terminate in the middle of a group would break
>some of their reordering mechanisms which kind of rely on branches always terminating a group.
I agree and I doubt that they would have modified their internals too much from the previous iterations. However the other constraints of group bundling (like FP ops positions, etc...) might have changed. Potentially becoming even more strict to overcome the shorter cycle time unless they have done something magical to the front-end ;).
---------------------------
>I suspect that branch slots must terminate a group in the POWER6. They might let
>you use the 5th slot for something else now though...
>
>I just think having basic blocks terminate in the middle of a group would break
>some of their reordering mechanisms which kind of rely on branches always terminating a group.
I agree and I doubt that they would have modified their internals too much from the previous iterations. However the other constraints of group bundling (like FP ops positions, etc...) might have changed. Potentially becoming even more strict to overcome the shorter cycle time unless they have done something magical to the front-end ;).
