[llvm] r225034 - InstCombine: try to transform A-B < 0 into A < B
kevinqindev at gmail.com
Wed Jan 14 18:36:13 PST 2015
Sorry that I didn't notice David had committed a patch to fix it. I'll look
at the performance results after applying that patch.
2015-01-15 10:01 GMT+08:00 Chandler Carruth <chandlerc at google.com>:
> On Wed, Jan 14, 2015 at 5:53 PM, Kevin Qin <kevinqindev at gmail.com> wrote:
>> Hi Michael and David,
>> I also see some performance regressions on aarch64 target introduced by
>> this commit. No further invesitgation on what's the root cause, but I guess
>> the one Michael shared may cause regression on many targets, including
>> I argee with Michael's suggestion that revert this commit firstly, and
>> make more performance test over it.
> It would be good to understand whether the change David made early today
> would have addressed these issues or not.
> (Either way, it has made me suspect that we may want A - B < 0 to be the
> canonical form, and lower it cleverly in the backend.)
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-commits