[PATCH] D104237: [RISCV][VP] Lower FP VP ISD nodes to RVV instructions

Thu Jun 17 02:28:43 PDT 2021

frasercrmck added a comment.

In D104237#2823780 <https://reviews.llvm.org/D104237#2823780>, @rogfer01 wrote:

> Thanks for looking into it. In our downstream we defer it too to a function call so we can complete the compilation and elsewhere we can provide an implementation (there is a number of other ISD nodes that we already lower this way like `exp`, `sin`, `cos`, etc.). AFAICT Arm SVE is not lowering any operation to `LibCall`yet so perhaps this conversation should be raised in the mailing list. Not sure if we need to align `compiler-rt`/`libgcc` in terms of naming conventions here.
>
> I'm happy to land this without `frem`.

Thanks for the review.

Downstream for us is OpenCL and we have both a custom builtins/math library and whole-function vectorizer to transform `fmod` OpenCL functions (both scalar and fixed vector) into (wider & possibly scalable) vector forms. So we're currently just transforming any LLVM IR `fmod` into a call to the equivalent OpenCL builtin to avoid the backend complaining. It's not where I'd like to leave it but it does the job in the short term.

Yes perhaps we should arrange a call with SVE and any other interested parties. We just missed the opportunity to bring this up in the last SVE call for a while :)

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D104237/new/

https://reviews.llvm.org/D104237