lukel97 wrote:

I checked with all the patches in this series combined and I'm still seeing a 
12% regression with 538.imagick_r on RISC-V. The regression comes from no 
longer SLP vectorizing the hottest block, since LoopUnroll now runs before it.

I'm not sure why unrolling the loop prevents SLP from vectorizing it. We could 
explore moving LoopUnroll after SLP. Or alternatively look into getting SLP to 
kick in for the unrolled form.

https://github.com/llvm/llvm-project/pull/195886
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to