andykaylor wrote:

> To the extend of supporting the same llvm output as OG, are you suggesting we 
> should emit different LLVM? Or that OG also needs to be fixed?

Regardless of what we do here, I'd like to see the handling of vector-3 loads 
and stores clarified in the LLVM Language Reference. As for what we do in CIR, 
I was thinking maybe we could initially annotate the vector-3 loads and stores 
with some attribute indicating that a vector-4 load and store with a poison 
element is permitted/expected. If a similar extension were done in LLVM IR, 
this would lower naturally to a masked load/store on targets that support that, 
but assuming there will at least be a time before that happens in LLVM IR, I 
would say we could transform such an operation to the form proposed in this PR 
during LoweringPrepare.



https://github.com/llvm/llvm-project/pull/161232
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to