[llvm] [LV, VP]VP intrinsics support for the Loop Vectorizer (PR #76172)

Alexey Bataev via llvm-commits llvm-commits at lists.llvm.org
Wed Jan 31 08:58:07 PST 2024


alexey-bataev wrote:

> > Instruction::Add still does not work, crashes the compiler because this VPInstruction returns that it "does not use only first lane".
> 
> Ah yes thanks for checking, had a look at this and there's also another issue, that is how some VPInstructions access and store scalar values as `vector` values in `VPTransformState`, which doesn't interact well with other recipes in general. I'll soon share patches to address those 2 shortcomings, but I'd expect that they will take some time to land. Don't think we need to block this patch until that is done, replacing the custom opcode with `::Add` can be done when they land.
> 
> The latest version of the patch has diverged quite a bit from the description I think, could you update the title & description to reflect the current scope of the patch?

I checked the description, did not find big difference between the description and the code. Could you point explicitly, which part must be updated?

https://github.com/llvm/llvm-project/pull/76172


More information about the llvm-commits mailing list