[PATCH] D25722: Improved cost model for FDIV and FSQRT

Michael Kuperstein via llvm-commits llvm-commits at lists.llvm.org
Tue Oct 25 10:27:01 PDT 2016

mkuper added a comment.

In https://reviews.llvm.org/D25722#578473, @RKSimon wrote:

> I'm swaying toward using Agner's numbers for SandyBridge FDIV / FSQRT costs - would anyone have any objections?

I'd really like to understand the discrepancy between Agner's numbers and the IACA numbers - especially since IACA does give the expected number for the latency.
Andrew, any chance you could ping the IACA people at Intel and ask them about this?

Meanwhile, I think I'd also prefer this to go in with Agner's numbers - and change it to the IACA numbers later if it turns out it's accounting for something real.


More information about the llvm-commits mailing list