Draft patch for this. As a preparation for the upcoming benchmarking,
It's ok just be able to emit.
And after that,we will give a final design.
Thanks very much!
I updated the diff as I did some tests basing on this diff. And I'm going to add comments
to the bug(bug 27107).
As I mentioned in the mail, I think this diff shouldn't be the final design for inverse sqrt
problem,but It would be ok just emitting rsqrt for upcoming benchmarking.
But even though, I'm not sure it's right. Especially I tried to support operations in f32 in
If you could give me some advice, it would be very appreciated. : )