rj-jesus wrote:

> Without the default cost-model committed this was looking pretty bad on both 
> Arm and AArch64, but with it it seemed OK overall. (There are lots of healthy 
> improvements). I don't think our Neon mulh optimization are as good as they 
> could be as they came up less frequently in DAG before this.

Yeah #224002 should be merged first (now done). Are you aware of any other 
tidy-ups we should also do? Unsigned mul overflow checks may be one 
(https://godbolt.org/z/Y1nY9Gb7v), and the AArch64-specific costs as you wrote 
in the other PR.

> Looks like SLPVectorizer is not ready :)

I think that failure was unrelated, but the recently added default costs could 
expose other issues. 

https://github.com/llvm/llvm-project/pull/224311
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to