[llvm] r208255 - [X86TTI] Remove the unrolling branch limits

[Chandler Carruth]

On Wed, May 7, 2014 at 5:31 PM, Nadav Rotem <nrotem at apple.com> wrote:

> I think that removing the branch count limit make sense.  Do you mind
> sharing the performance data?  Were there any regressions or performance
> wins?
>

I saw no changes on sandybridge (or any other Intel or AMD processor) when
experimentally disabling this threshold. Neither regressions nor
improvements.


>
> I am asking about the performance data because I am guessing that there
> were some benchmarks that benefited from this heuristics otherwise it
> wouldn’t have made it in.
>

It wasn't based on data, it was based on a reading of the relevant
optimization guidelines. However, those guidelines are actually talking
about *taken* branches, and we have no terribly easy way to model that.
Also, it turns out to be very hard to craft a case where you would exceed
the number of taken branches and yet stay under the basic threshold, so it
seems like the basic one is the only one needed.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140511/320cbb1b/attachment.html>