[PATCH] blockfreq: Rewrite block frequency analysis

Chandler Carruth chandlerc at google.com
Wed Mar 19 13:45:56 PDT 2014


On Wed, Mar 19, 2014 at 1:38 PM, Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:

> Does this sound good to everyone?
>

Well, I already said the basic cleanup patches were good.

I've not really had a chance to look at the algorithm in detail yet because
...


>
> Chandler, do you still want something up on Phab at some point?
>

... I was waiting for this before reviewing the algorithm in detail, so I
still haven't had a chance to do that. If using phab is slow, I can just go
through the patch files, I don't really care, it just seemed like i should
wait based on prior emails.


Currently, I'm somewhat skeptical of the benefit of PositiveFloat *for this
use case* (not in general), largely because of the comments about using
power-of-two loop trip scales. I actually don't understand why
distinguishing between a scale of 3 and 5 would be interesting at all. Do
we have any reason to believe either than our profiles will be this
accurate or that we will have transformations which are this sensitive? I
don't think either are true.


 If so, does it
> makes sense for me to post the current, massive, combined patch of
> 0003+0004+0005+0006(+0007?) now, so you can look at the algorithm as a
> whole?
>

I'm actually interested in evaluating this totally independent of the bias
part. I'm in complete agreement that the flow-solve aspect is orthogonal to
what precise metric is useful for triggering different optimizations.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140319/dfbaf995/attachment.html>


More information about the llvm-commits mailing list