davemgreen wrote:

The MVE side is now all committed, making sure that we don't scalarize the 
intrinsics. There was a question about whether we should be using target 
specific intrinsics for it, but I believe we need to as we can't really convert 
the constrained intrinsics into mve instructions. I think we need something 
similar for Neon, at least to stop the intrinsics from scalarizing. I've not 
looked into how many different types of nodes that would be.

A few other things I have noticed:
 - ldexp does not seem to work for fp16 vectors 
https://godbolt.org/z/MW3jqx797. It should be able to scalarize.
 - There are a number of problems with the various min/max's, some that are 
present for both standard nodes and strictfp.
 - On systems without fullfp16, fp16 nodes do not lower at the moment 
https://godbolt.org/z/3qh3fbWjG. (For systems without fp64 we do manage to 
produce libcalls).

https://github.com/llvm/llvm-project/pull/137101
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to