[PATCH] D26855: New unsafe-fp-math implementation for X86 target

[Chandler Carruth via Phabricator via llvm-commits]

chandlerc added a comment.

In https://reviews.llvm.org/D26855#650115, @chandlerc wrote:

> Unless the change is trivially and obviously an improvement on inspection of the result, I think you need data before making it. =] I've not looked at this particular change in enough detail to say whether it satisfies that. Maybe Gerolf or Craig could.


Another way besides benchmark data to show runtime improvements are tools like IACA. This is how we curated most of the vector shuffle improvements over the last few years for example.


https://reviews.llvm.org/D26855