[PATCH 3/3] [PBQP] Enforce a safe order for ConservativelyAllocatableNodes

[Lang Hames]

Hello Arnaud,

This is a good idea, but the implementation in this patch will make removal
of conservatively allocatable nodes more expensive: from O(N) to O(RN^2),
where R is the number of registers and N is the number of nodes. We can get
that back down to O(RN + NlogN) by creating a VectorMetadata class that
contains the expected/worst-case cost, and using a set (or better yet, a
priority_queue) to hold the conservatively allocatable nodes. This is the
same line of reasoning that was applied (very successfully) to
MatrixMetadata to bring down the cost for the conservative-allocatability
test itself.

If you're happy to rewrite your patch along those lines please go for it,
otherwise I can find some time to do it in the next few days. It is not a
big change at all, though it helps that I'm familiar with the
infrastructure.

Cheers,
Lang.


On Wed, Oct 22, 2014 at 6:43 AM, Arnaud A. de Grandmaison <
arnaud.degrandmaison at arm.com> wrote:

> Hi Lang,
>
>
>
> Here is the last patch in the PBQP improvement serie. Could you please
> have a look at it ?
>
>
>
> Nodes in ConservativelyAllocatableNodes were processed in the order they
>
> have been added in NodeSet; however, it is preferable to process nodes in
>
> an order based on their cost, to ensure we keep the total cost under
>
> control and provide a reasonable solution.
>
>
>
> Cheers,
>
> --
>
> Arnaud A. de Grandmaison
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141104/bc7381a4/attachment.html>